Investors spend enormous amounts of time debating the next Nvidia architecture , the next capacity announcement and the next financing package. The cash earned by hardware already installed deserves the same attention.

CoreWeave offers a concrete example. On its second-quarter earnings call, management disclosed an A100 contract extending into 2029. That architecture debuted in 2020. Management also described prior-generation pricing as at or above earlier levels. [1]

This is evidence of a market finding productive uses for older compute. It does not establish that every A100 has been running continuously since 2020, or that every cluster can generate attractive returns for nine years. But it challenges a blanket assumption that a new architecture rapidly eliminates the earning power of its predecessors.

My bullish interpretation is straightforward: profitable renewals can turn the installed fleet into a much longer stream of cash flows than a short obsolescence assumption allows. The stock implication depends on the price paid for that stream and the financing required to build it.

Subscribe now

The evidence is becoming harder to dismiss

Nvidia’s October 1 blog organizes AI-factory returns around earning capacity, useful life and demand. Its emphasis on durability is particularly relevant: a system’s ability to serve different workloads can keep it commercially useful after a new generation ships. The company points to A100 bookings through 2029 as an example. [2]

Its September investor presentation adds another dimension. Using Silicon Data monthly residual-value estimates, Nvidia shows A100 at 25% of assumed original cost, H100 at 58% and B200 at 158%. The systems are different ages. The cost assumptions are $20,000, $35,000 and $45,000 per GPU respectively, on an eight-GPU HGX-system cost allocation. [3]

That changes the precision of the appreciation argument. For B200, the presentation supports an estimated value above assumed acquisition cost. For A100 and H100, the evidence supports substantial retained value. Those are different outcomes, and both matter.

The supplied a16z chart provides a separate look at rental earning power: its September endpoints are $1.59 per GPU-hour for A100, $2.63 for H100, $3.29 for H200 and $5.69 for B200. Its accompanying research describes resilience in older-generation rental pricing. [4]

Rental rates help reveal what customers will pay to use an asset. Resale estimates help reveal what someone might pay to own it. Neither replaces the operating-cost and financing analysis needed to calculate shareholder returns.

The frontier moves; workloads spread across generations

On TBPN, Mike Intrator described getting “clubbed in the head” over obsolescence concerns and pointed to an “incredibly fat tail” of productive uses beyond frontier workloads. [7] His argument offers a useful way to think about this market. Frontier workloads place a premium on the latest systems. Other customers prioritize a familiar software stack, an existing deployment, availability or a lower total bill for a task with less demanding performance requirements.

The demand curve can therefore broaden across generations. An older cluster can keep serving useful work while a newer cluster expands what is possible at the frontier. That is the hypothesis underlying the durable-renewal case, not a claim that older hardware always offers better economics.

Nvidia’s software argument adds another potential source of durability: improvements to libraries and kernels can increase the usefulness of an installed system without replacing its silicon. [2] The investment inference is that hardware age alone may be an incomplete description of the asset’s earning capacity.

This is where the compute paradox becomes relevant. Falling cost per unit of intelligence can make additional applications economical. If that expands demand enough, multiple generations may remain commercially occupied. The evidence so far supports testing that possibility; it does not establish unlimited demand or permanently rising prices.

The economic value sits in the later contracts

The first contract has to justify the original investment. It must support operating costs and capital recovery, while the financing schedule determines how much cash is available to equity holders.

A subsequent contract can be valuable because it reuses equipment, network integration and an operational deployment. The original capex is already sunk. Continued service still requires maintenance, power, rent, support and sometimes incremental refurbishment—but it may avoid the large upfront expenditure associated with an entirely new cluster.

That creates a distinction investors should model explicitly:

Accounting depreciation allocates historical cost over an estimated useful life. Economic life describes how long an asset can earn attractive returns. Market value reflects what a buyer might pay today.

A longer depreciation schedule can raise reported earnings without creating operating cash. A longer profitable earning life adds future cash flows. Higher sale proceeds add value when the asset is actually disposed of. A sound model keeps these effects separate.

The depreciation chart also requires nuance. Operators have extended useful-life estimates over time, but those estimates do not move in only one direction. Amazon shortened the life of a subset of servers and networking equipment from six years to five beginning in 2025, citing faster technological development. [5] Broad server schedules are useful context, not a universal GPU-lifetime rule.

What two additional earning years are worth

To isolate the lifetime effect, I held the other assumptions constant in an illustrative project model:

• $40,000 of initial installed cost per representative GPU-equivalent.

• $3.26 per billed GPU-hour and an 85% commercial billing factor.

• Identical pricing for the first three years, followed by a 5% first-renewal reduction and 5% annual reductions thereafter.

• A 63% cash operating margin, $1,000 of first-year maintenance growing 3% annually, and explicit cash taxes.

• A six-year straight-line tax-depreciation proxy, 5% terminal resale and a 12% unlevered discount rate.

These are analyst assumptions, not disclosed CoreWeave unit economics. [6]

The four-year project produces approximately negative $2,358 of NPV. Six earning years produce positive $8,551. Eight produce positive $15,170.

Moving from six years to eight adds approximately $6,619 of present value per representative installed GPU-equivalent, even while rental rates decline. This is the key result: durable economics can improve lifetime returns without requiring perpetual appreciation.

The model separately tests more aggressive price erosion and stronger renewals. Its sensitivity grid makes the relationship visible.

A longer life creates room for weaker pricing. It does not rescue every project. With enough price erosion, operating cost or upfront investment, extending the calendar can simply extend poor economics.

The equity upside runs through capital productivity

The most interesting bullish scenario is a business in which new capacity drives growth while older capacity keeps contributing cash.

If that happens, CoreWeave could earn more lifetime cash per dollar invested. Mature clusters could help fund expansion, support debt repayment and reduce the need for incremental equity financing. Evidence of repeatable capital recovery could also improve investors’ confidence in the durability of returns.

Each step is conditional. Growing EBITDA does not equal free cash flow, and a valuable installed fleet does not eliminate the cost of the next buildout. Higher collateral estimates may improve financing flexibility, but additional borrowings remain liabilities.

The valuation work therefore needs a sequence: deployment dates, realized billing, after-rent margins, maintenance and replacement expenditure, cash taxes, financing and dilution. Only then does an enterprise-value estimate become a common-equity value.

I would also resist multiplying the representative-project NPV by the projected year-end GPU count. Actual fleets contain different architectures, acquisition costs, contract terms and ages. A GPU added in December contributes little to that calendar year’s revenue. The corrected model distinguishes average deployed capacity from year-end capacity and keeps project economics separate from the corporate valuation bridge. [6]

The risk that can still defeat the thesis

Power is an opportunity cost. An older cluster can remain profitable in isolation and still be a less attractive use of scarce energized capacity than a newer system. Redeployment, infrastructure compatibility and the cost of conversion determine whether that comparison is actionable.

The other constraints are equally concrete: maintenance can rise, renewals can weaken, customers can fail to pay and expansion can require more financing than expected. A surviving chip is not automatically an attractive investment.

The question I want answered over the next several earnings reports is how much capital each mature cluster has recovered—and how much additional cash it can earn after renewal. Architecture-level renewal pricing, after-rent contribution margins, maintenance needs and financing amortization would make that answer much more useful than a single fleet-wide useful-life estimate.

For now, the A100 contract and residual-value estimates make a longer earning tail a serious underwriting case. The opportunity is to identify which operators can convert that durability into cash available to shareholders.

The next architecture expands the frontier. The next contract reveals the value of the installed fleet.

Sources and methodology

[1] CoreWeave, Q2 2026 earnings call, August 11, 2026. Management statements on prior-generation pricing and A100 bookings through 2029

[2] Nvidia, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” October 1, 2026

[3] Nvidia, September 2026 investor roadshow, slide 22. Underlying monthly residual-value estimates credited to Silicon Data

[4] a16z, “State of Markets II,” and user-supplied rental-price graphic credited to Silicon Data, dated September 7, 2026

[5] Amazon, Q2 2025 Form 10-Q, useful-life estimate change

[7] TBPN, interview with Mike Intrator, September 30, 2026, episode “White House AI Accord, Model Welfare, Ken Griffindor School of Market Wizardry”

Disclaimer: Research and scenario analysis, not a verified fair-value conclusion. Nvidia and CoreWeave have commercial interests in the durability argument; reported management claims and estimated market values should be tested against realized cluster cash flows.

Subscribe now

Share LR Market Intelligence

That comparison is directionally right, but financially incomplete.

Subscribe now

Oil is a commodity. Compute is a productive asset. A barrel disappears when it is consumed; a GPU can be rented repeatedly. The more useful comparison is therefore not between a GPU and a barrel of oil. It is between a GPU and the well that produces it.

That distinction became more interesting when Ornn Exchange introduced its B300 Payback Forecast. Under the assumptions displayed: 80% utilization and a forward curve for rental prices, the model estimates that an NVIDIA B300 can recover its purchase price through rental revenue in approximately 14 months.

That is not a guaranteed return. It is a modeled payback period whose outcome depends on purchase cost, utilization, rental pricing, operating expenses and the eventual residual value of the hardware.

But if the estimate is even approximately correct, its message is significant:

A newly deployed B300 may possess the payout economics of a high-quality shale well.

The comparison also gives us a better framework for underwriting the emerging neocloud industry.

The B300 as a digital oil well

In oil and gas, an operator commits capital to drill and complete a well. Once production begins, the first question is payout: how long will it take the well’s cumulative cash flow to recover the original investment?

The same logic applies to compute.

A neocloud purchases GPUs, connects them to power and networking, installs them inside a functioning cluster and sells access by the GPU-hour. Once the cluster becomes available to customers, it begins producing revenue.

In Landman terms:

The B300 is powerful hardware. NVIDIA’s DGX B300 configuration includes eight Blackwell Ultra GPUs, 2.1 terabytes of total GPU memory and 14.4 terabytes per second of aggregate NVLink bandwidth. NVIDIA positions the architecture specifically for increasingly compute-intensive reasoning and inference workloads. NVIDIA DGX B300 specifications

But technical performance alone does not establish investment value.

What matters is how efficiently that performance can be converted into cash.

Is 14 months good?

At the hardware level, a 14-month payback is excellent.

The asset theoretically returns its initial capital in approximately 1.17 years. If it then remains economically productive for several additional years, the cash generated after payout can produce an attractive return on invested capital.

In oilfield language, that resembles a good well on Tier‑1 acreage:

• Capital is deployed quickly.

• Production begins relatively quickly.

• Early cash generation is strong.

• The initial investment is recovered before the asset reaches its mature decline phase.

• Post-payout cash can fund the next development cycle.

This is precisely why payback matters in a rapidly changing technology market. The operator does not need the B300 to remain the world’s leading GPU forever. It needs the chip to return its capital before superior hardware materially erodes its rental economics.

That is the compute equivalent of getting a shale well to payout before its steepest decline overwhelms the initial production.

Yet the comparison requires an important qualification.

A 14-month GPU payback is not necessarily a 14-month neocloud payback.

The well is not the field

Ornn’s estimate focuses on the chip’s purchase price and expected rental income. A publicly traded neocloud owns a much larger economic system.

Its complete investment may include:

• Servers and networking equipment

• Electrical infrastructure

• Cooling systems

• Data-center construction or leases

• Software and orchestration

• Customer-acquisition expenses

• Corporate overhead

• Interest expense

• Financing fees

• Deployment delays

• Equity dilution

In the oil patch, analyzing only the well’s drilling cost while ignoring acreage, gathering infrastructure, transportation, overhead and financing would be a form of half-cycle underwriting.

The same mistake can occur in compute.

A B300 might be an exceptional well while the company operating it remains a mediocre investment. Management can still destroy value by paying too much for power, accepting punitive financing, building speculative capacity, missing deployment schedules or allowing hardware to sit idle.

That creates four distinct levels of analysis:

This is the most important distinction for investors:

GPU payback measures asset quality. Equity returns measure asset quality after execution, overhead and capital structure.

A great well can still sit inside a bad E&P.

The neocloud payback equation

A simple GPU payback calculation begins with:

Where:

• U = realized utilization

• R = realized rental revenue per GPU-hour

• V = variable cost per available GPU-hour

• 8,760 = hours in one year

The result is expressed in years. Multiplying by 12 converts it into months.

This formulation improves upon a simple purchase-price-versus-rental-revenue calculation because it uses the hourly contribution margin rather than gross rental revenue.

It also reveals exactly what can break the economics.

Installed GPU cost

The numerator should include more than NVIDIA’s invoice price. A more conservative calculation incorporates the GPU’s proportional share of:

• Host server equipment

• High-speed networking

• Installation

• Power distribution

• Cooling infrastructure

• Initial deployment expense

A cheap chip inside an expensive cluster may not actually possess a short full-cycle payback.

Utilization

Utilization is compute’s production volume.

An oil well may underperform because geology disappoints. A GPU underperforms when it sits idle, experiences deployment delays or cannot attract sufficiently valuable workloads.

An 80% utilization assumption means the GPU generates billable revenue for approximately 7,008 hours annually. At 50% utilization, that falls to 4,380 hours.

Because utilization sits in the denominator, modest changes can create large payback extensions.

Holding everything else constant:

These figures are sensitivity calculations, not Ornn forecasts.

Rental rate

Rental pricing is compute’s realized commodity price.

Ornn’s OCPI methodology uses transacted rental prices rather than advertised cloud listings. Its public market page currently shows separate settled prices for H100, H200, B200 and A100 systems, illustrating how each generation develops its own rental market. Ornn OCPI market data

The B300’s forward curve is therefore crucial. Today’s scarcity price is less important than the average realized price earned throughout the payback window.

Variable cost

Power is compute’s lifting cost—but it is not the only one.

Variable and semi-variable expenses can include:

• Electricity

• Cooling

• Data-center fees

• Network transit

• Maintenance

• Support

• Revenue sharing

• Workload-orchestration expense

The operator with the lowest all-in cost can continue producing profitably after higher-cost competitors become uneconomic.

That is why secured low-cost power resembles Tier‑1 acreage.

What happens when assumptions deteriorate?

Starting from a 14-month baseline, we can stress the two most important commercial variables: utilization and hourly contribution margin.

The table below applies proportional changes to the original economics. “Contribution-margin compression” means the net amount earned per utilized GPU-hour declines after variable costs.

The lesson is not that 14 months is unreliable. The lesson is that the distribution matters more than the headline.

At 80% utilization with durable pricing, the B300 resembles a Tier‑1 well.

At 50% utilization with 40% contribution-margin compression, the same asset takes more than three years to pay out. That creates a far narrower margin before newer hardware arrives.

Compute’s decline curve

Shale wells and GPUs both generate front-loaded economics, but their decline mechanisms differ.

A shale well’s physical production typically falls quickly after initial production. EIA data show why shale operators must continuously add new wells to offset declines from existing production. In 2024, more than 15,000 new Lower-48 wells were required to counteract substantial declines from older wells. U.S. Energy Information Administration

A GPU can continue operating at approximately the same physical capacity. Its economic output declines differently:

1. New supply pressures rental rates.

2. A new architecture improves performance per dollar.

3. Premium workloads migrate toward newer chips.

4. Older hardware experiences declining utilization.

5. The asset cascades into lower-value inference or secondary workloads.

Oil’s decline is principally geological.

Compute’s decline is competitive.

That makes the B300 less like a conventional field with decades of predictable production and more like shale: high initial productivity, attractive early payout and a constant requirement to replenish the asset base.

The compute Red Queen

The best neoclouds can create a reinvestment flywheel:

1. Deploy the newest GPUs.

2. Capture early scarcity pricing.

3. Maintain high utilization.

4. Recover invested capital quickly.

5. Use the cash flow and collateral value to finance the next generation.

6. Repeat before existing hardware loses premium status.

This sounds attractive, and it is, but it also creates a Red Queen problem.

The operator must keep running just to maintain its competitive position.

If the company stops reinvesting, its fleet ages. If it reinvests too aggressively, it may produce revenue growth without producing free cash flow. If it uses expensive debt to maintain the cycle, creditors may capture more of the economics than shareholders.

This is why rapid growth alone is insufficient evidence of value creation.

The central question is:

Is reinvestment compounding the equity, or merely preventing the fleet from becoming obsolete?

The neocloud acreage map

The oilfield analogy also helps separate high-quality neocloud operators from promotional ones.

Tier‑1 compute acreage possesses five attributes:

1. Power

The operator controls sufficient, affordable and deliverable electricity.

Announced megawatts are not the same as energized megawatts. Energized megawatts are not automatically revenue-generating megawatts.

2. Deployment speed

The operator can transform power, buildings and GPUs into usable capacity quickly.

A B300 sitting in inventory is the equivalent of a drilled-but-uncompleted well: capital has been committed, but production has not begun.

3. Contract quality

High utilization is far more valuable when supported by durable customer commitments instead of volatile spot-market demand.

Investors should distinguish between:

• Contracted and deployed

• Contracted but undelivered

• Available and uncontracted

• Announced but unfunded

4. Cost of capital

Two operators can own identical GPUs and generate radically different equity returns if one finances them at a materially lower cost.

In capital-intensive businesses, funding is part of the operating model.

5. Residual-value strategy

Older GPUs may retain economic usefulness through inference, fine-tuning, research, smaller models and lower-priced cloud tiers.

The best operator does not merely buy the newest hardware. It manages the entire fleet’s workload cascade.

What the 14-month estimate actually tells us

The Ornn forecast does not prove that every B300 deployment will generate an extraordinary return. Nor does it prove that every neocloud deserves a premium valuation.

It tells us something more useful:

Current compute economics may permit new hardware to recover capital before technological obsolescence becomes fatal.

That supports several bullish conclusions:

• Advanced compute remains scarce.

• Customers are still paying substantial premiums for frontier hardware.

• GPU fleets can support asset-backed financing.

• Rapid payout can create internally reinforcing capital cycles.

• Efficient operators may compound deployment faster than slower competitors.

• Low-cost power and contracted utilization are becoming strategic natural resources.

But it also raises a harder question for public-equity investors.

If GPU-level economics are this attractive, capital will chase them. New supply will enter. Rental rates will compress. The advantage will migrate away from simple GPU ownership and toward the operator with the best power, deployment, customer and financing machine.

That is what happened in shale.

The first advantage was access to the resource. The enduring advantage became the ability to develop it more efficiently than everyone else.

Conclusion: compute is the new shale

The B300’s modeled 14-month payback should not be interpreted as a guaranteed clock.

It should be interpreted as an economic signal.

At 80% utilization and along Ornn’s expected rental curve, the newest generation of NVIDIA compute may possess the rapid payout characteristics of a high-quality shale well. That makes it highly financeable, highly productive and potentially capable of generating substantial post-payout cash.

But a chip is not a cluster. A cluster is not a platform. And a platform is not the equity.

The winning neocloud will not simply own the most GPUs. It will control the best acreage, power, deployment capacity, customers and capital, and convert that position into durable cash returns before the technology turns.

Oil companies drill to replace declining production.

Neoclouds deploy to replace declining economic relevance.

Both operate on a treadmill. Both can generate exceptional returns. And in both industries, the winners are not necessarily those who spend the most capital.

They are the ones who recover it fastest, and reinvest it best.

Disclaimer: This is for Educational purposes only. NFA.

Subscribe now

Habitat, its application storage platform, handles more than 70 million requests per second. Two engineers, working with Codex and GPT-5.5, rewrote the service in Rust. OpenAI reports 6x CPU efficiency and 15x memory efficiency versus Python. These are service-level results, not reductions in its total infrastructure bill. OpenAI, September 11.

The economic possibility is compelling: AI tools can help improve the software that determines how much computing a business needs to operate.

That creates a second way to participate in the AI buildout. Investors can examine the businesses buying infrastructure and ask how much more profit those assets might support.

My preferred public-equity candidate for that question is Meta . The thesis depends on whether productivity gains survive reinvestment and reach shareholders.

Subscribe now

What the engineering means

Habitat sits between products and their stored information, handling routing, access controls and related work. In the supplied diagrams, responses can wait even after the database is ready because CPU work is blocking their processing. Other diagrams show connection reuse directing more traffic toward already overloaded servers.

Think of a distribution center where packages have arrived but cannot leave the loading dock. Adding warehouse space does little for that particular bottleneck. Better scheduling and faster handling can increase useful throughput from existing assets.

For investors, that distinction matters. An improvement inside a software service does not tell us how many chips its operator will buy next year. The company might use the capacity it frees to serve more demand, improve its product, delay purchases, or reduce costs. Each choice produces a different result for suppliers and shareholders.

The investment work begins with identifying which choice management is likely to make.

Follow the savings to the income statement

My working hypothesis is that the strongest beneficiaries will be businesses whose revenue is tied to a valuable customer outcome while their delivery costs fall.

Consider three hypothetical business models.

1. An advertising platform improves its systems and needs less computing to deliver the same advertising results. It may retain those savings without automatically cutting advertising prices.

2. A cloud vendor improves its infrastructure. It can gain margin or serve more demand, but customers optimizing their own applications may consume fewer billable resources. The result depends on pricing, capacity constraints and workload growth.

3. A software services firm completes a migration with fewer developer hours. Under hourly billing, revenue per project may decline. Under a fixed-price contract, the same productivity gain can improve margin. Lower prices may also unlock projects clients previously could not justify.

Productivity does not identify the winner by itself. The contract and the revenue model determine who captures it.

This is why I would not automatically turn the OpenAI article into a short on infrastructure. Nor would I automatically buy every company that uses AI coding tools. The relevant question is whether the benefit remains inside the business after customers, competition and reinvestment take their share.

Why Meta is the candidate

Meta offers an attractive structure for retaining savings: advertising monetization paired with a large internal technology operation. Advertisers are buying access to audiences and advertising performance; they are not purchasing a fixed quantity of Meta’s CPU time.

That opens two possible routes to better economics. Meta could deliver an existing level of service more cheaply. It could also redirect productivity into better ranking, measurement, product development and advertising tools, potentially increasing revenue from the same audience. The second route requires evidence of monetization; it cannot simply be counted as cost savings as well.

Its latest results show why this is worth studying, and why the conclusion must remain conditional.

In Q2 2026, revenue rose 28%, while total expenses increased 55% and operating income fell 8%. Expenses included $2.40 billion in legal charges and $1.18 billion in severance. Capital expenditures, including finance-lease principal, reached $31.08 billion, leaving $784 million of free cash flow. Meta Q2 results.

The opportunity I want to test is that demand remains healthy while the cost required to support each additional dollar of revenue improves.

There is an important objection: Meta already runs mature infrastructure. We cannot transplant another company’s rewrite results onto its consolidated cost base. A genuine earnings thesis needs a plausible pool of addressable costs, a feasible improvement rate and a credible reason management will retain some of the benefit.

Until those inputs are established, Meta is my preferred candidate for the mechanism, rather than a demonstrated bargain at any price.

The earnings breakeven frontier

The useful chart is a map of what happens after an efficiency improvement.

Define annual recurring cost savings as S and added annual income-statement expense as I, both relative to the same baseline. Hold revenue and other items constant. With an assumed 16% tax rate and 2.566 billion diluted shares, the incremental annual earnings effect is:

ΔEPS = (S − I) × 0.84 ÷ 2.566, with S and I measured in billions of dollars.

The share-count assumption is anchored to Meta’s Q2 reported diluted weighted-average shares; this is a sensitivity model, not company guidance or a consensus forecast. Meta financial tables.

At point A, $5 billion of recurring savings accompanies $2 billion of added expense. The $3 billion net benefit contributes approximately $0.98 in annual EPS.

At point B, the same $5 billion of savings accompanies $7 billion of added expense. Annual EPS is approximately $0.65 below baseline. The efficiency gain still helps relative to spending the same amount without achieving it; it simply does not offset the entire spending increase.

The chart makes the decision visible. Moving right improves current earnings through savings. Moving up absorbs that benefit through additional expense. A company can become more technically efficient while moving into a worse near-term earnings position.

That does not necessarily mean it is destroying value. Spending above the line may produce future revenue and attractive returns. The diagonal is an earnings breakeven frontier, not a verdict on long-term investment quality.

It also is not a cash-flow model. Capital expenditure is a cash outlay; depreciation enters earnings over time. Meta could improve EPS and still absorb substantial cash in construction and equipment. Both statements must be underwritten.

Turn a mechanism into an expectation gap

Each hypothetical $1 billion of net annual pretax savings contributes about $0.33 of EPS under this model. At an illustrative unchanged 24-times earnings multiple, $5 billion of net savings would contribute about $39 per share of incremental earnings-based value. Neither the savings nor the multiple is a forecast or a current valuation claim.

That calculation puts discipline around the enthusiasm. Efficiency alone must be substantial to move the value of a company this large. The stronger case combines retained savings with sustained revenue growth and better returns on investment.

For the stock to surprise positively, the improvement also has to exceed what investors already expect. Savings already embedded in earnings estimates cannot be counted again as upside. The relevant S and I would need to measure deviations from those estimates, not merely changes from last year.

I have not established that current expectations omit this benefit. That is the remaining underwriting task, and the reason this article presents a candidate and a test rather than a price target.

Over the next two earnings reports and the first detailed 2027 spending outlook, I would look for three things: durable advertising demand; slower underlying expense growth after separating unusual charges; and better conversion of operating cash flow into free cash flow. I would also distinguish real operating improvements from changes in depreciation assumptions.

The thesis weakens if the addressable savings pool is small, benefits are continuously absorbed by spending without improving returns, or advertising demand deteriorates. A lower cost per task does not rescue weak monetization by itself.

Where the pressure could appear

The same framework puts labor-based software services on the other side of the research list. EPAM derived roughly 78% of Q2 revenue from time-and-materials contracts, calculated from its disclosed contract mix. That creates exposure to fewer billable hours per project. EPAM Q2 filing, via Quartr.

But lower delivery costs could expand demand, and fixed-price work can retain savings. I would want evidence of weakening project economics before treating the company as a short. The contrast is useful for understanding the mechanism; it is not sufficient to construct a pair trade.

The question I want to carry into the next earnings season is simple: how much of the productivity improvement becomes incremental profit, and how much gets spent again?

Meta is where I would investigate first. The opportunity would become compelling if strong advertising economics meet a cost trajectory that improves faster than expectations. That is the efficiency dividend worth underwriting.

Methodology: Company figures are reported results; indexed resource intensity and EPS sensitivities are calculations. Scenario inputs are explicitly hypothetical. No GPU-efficiency or company-wide savings estimate is inferred from Habitat. Sources were reviewed on September 11, 2026. Not financial advice. Not a solicitation to buy or sell securities. For educational purposes only.

The deeper I went, the more puzzled I became by how differently the market seemed to view the sector.

Eventually, I realized the disconnect was not necessarily that others were missing an obvious fact. It was that understanding the economics of these businesses requires following several layers of the AI stack at once. And increasingly, one of the most important layers may be the newest and least understood of them all: the economics of tokens.

Token economics are still developing in real time. We do not yet know exactly how the long-run relationship between falling inference costs, exploding usage, model efficiency, and physical compute demand will settle.

But after months of following the data, I kept hearing the same signal.

There was something about the neocloud model, and CoreWeave in particular, that I could see in the numbers but had struggled to articulate clearly. The more I studied it, the more the pieces began to fit together.

This piece is my attempt to finally put that idea into words.

My goal is to explain why I believe the value proposition of the neoclouds extends far beyond the two narratives that dominate the debate today:
“Compute is scarce.” and “There is too much leverage.”

Both matter. Neither, in my view, captures the full economics of what is happening.
Because there is a strange divergence developing inside the AI economy.
And I’ve decided to call it The Compute Paradox.

The Compute Paradox: As the cost of intelligence falls, demand for it can rise even faster, causing total compute consumption, and potentially the value of the infrastructure supplying it, to increase rather than decline.

When Cheaper Intelligence Demands More Compute

There is a strange divergence developing inside the AI economy.
The price of intelligence is collapsing.
The price of the compute required to produce it is not.

That distinction may end up being one of the most important signals in the entire AI infrastructure trade.

Citadel Securities recently published Elastic Expectations, an excellent follow-up to its earlier work on Tokenomics. The central question is deceptively simple:

When intelligence becomes cheaper, do users spend less on AI, or do they consume so much more intelligence that aggregate spending actually increases?

The latest evidence increasingly points toward the second outcome.

Usage-weighted token prices have fallen roughly 40% since the end of June, according to Silicon Data. Yet Citadel calculates that July AI spend per employee increased approximately 49% month over month among the top 1% of enterprises, 25% among the top 10%, and 9% at the median. GPU rental prices, meanwhile, have rebounded from their June lows even as token prices continued falling.

This is not what the simple AI commoditization bear case was supposed to look like.

And for companies such as CoreWeave, it may fundamentally change the valuation framework.

The Economics of Cheaper Intelligence

The conventional AI infrastructure bear case is straightforward:

Better GPUs → greater efficiency → cheaper inference → lower compute prices → commoditization → lower infrastructure returns

There is nothing irrational about that argument.

In fact, if demand for intelligence were relatively inelastic, it would probably be correct.

But there is another possibility.

Better GPUs → cheaper intelligence → more economically viable AI workloads → exponentially more consumption → higher aggregate compute demand

This is essentially Jevons paradox applied to intelligence.

Efficiency does not necessarily destroy demand for a resource. Sometimes it makes the resource useful enough that total consumption increases.

We may now be watching that happen in real time.

The more affordable intelligence becomes, the more intelligence the economy may decide to consume.

That distinction matters enormously.

Citadel’s most important chart may be the divergence between token prices and H100 rental prices.

Earlier this summer, the two were declining together.

That was dangerous.

Falling token prices alongside falling compute prices could have indicated something much worse than technological efficiency: weakening demand across the entire AI stack.

But that relationship has since broken.

GPU rental prices measured by Ornn have rebounded while token prices have continued to decline. Citadel describes this as tentative evidence that the Jevons effect may be becoming more dominant. Lower-priced intelligence appears to be stimulating enough additional usage to support demand for the underlying physical compute.

That is an extraordinarily important development for the neoclouds.

Quantity Is Exploding

Now add the quantity side.

OpenRouter recently crossed approximately 75 trillion tokens per week, with the subsequent week pacing above 87 trillion tokens.

That represents an annualized run rate north of 4.5 quadrillion tokens.

OpenRouter’s chart is remarkable because it is presented on a logarithmic scale and still looks almost like a straight line.

From roughly:

• 10 billion tokens per week in January 2024

• 1 trillion per week by March 2025

• 10 trillion per week by February 2026

• 75 trillion-plus today.

The exact growth rate will obviously decelerate. No serious valuation model should extrapolate this curve indefinitely.

That is not the point.

The point is that the quantity response to declining intelligence costs is enormous.

Token prices can fall dramatically while the economic value of the infrastructure layer rises if quantity grows faster than price declines.

At the simplest level:

Revenue ≈ Price × Quantity

But AI infrastructure requires another layer:

Compute demand ≈ AI workloads × compute required per workload

The variable that matters for infrastructure investors is therefore not the nominal price of a token.

It is the total physical computation demanded by the economy.

The Compute Elasticity Ratio

While doing my research, I asked ChatGPT if there was a useful way to conceptualize this dynamic. What it came up with was actually pretty great.

Call it the Compute Elasticity Ratio, or CER:

CER = (1 + AI workload growth) ÷ (1 + effective compute-efficiency growth)

The intuition is simple.

CER < 1

• Efficiency grows faster than workloads.

• Physical compute demand eventually contracts.

• Bearish neocloud regime.

CER ≈ 1

• Additional workloads roughly absorb efficiency gains.

• Physical compute consumption stabilizes.

• Commodity regime.

CER > 1

• AI workloads grow faster than compute efficiency.

• Physical compute demand continues expanding despite relentless technological improvement.

• Bullish infrastructure regime.

And if CER remains materially above 1?

That is where the neocloud economics become extremely interesting.

The Firms That Understand AI Are Spending More, Not Less

Citadel’s enterprise spending data make this story even stronger.

The highest-spending AI companies are not showing signs of saturation.

They are accelerating.

Since October 2023, monthly AI spending per employee among the top 1% of firms has increased by approximately $6,542, compared with only $9.63 at the median. The ratio between 90th-percentile and median spending has roughly doubled from 27x to 54x.

Think about what this implies.

The companies furthest along the AI adoption curve are not discovering that they have “enough AI.”

They are discovering more things to do with it.

• Coding agents.

• Reasoning models.

• Long-context workloads.

• Multi-agent systems.

• Inference at scale.

• Video generation.

• Autonomous workflows.

• Scientific computing.

• Persistent enterprise agents.

The firms that have already learned how to consume intelligence appear to want dramatically more of it when its effective price declines.

This is exactly the demand curve an infrastructure owner would want.

Enter CoreWeave

This brings us to CoreWeave .

The common bear framework for CRWV views the company as a highly levered owner of rapidly depreciating GPUs selling a product that will eventually become commoditized.

That is the risk.

But it may not be the whole story.

According to its Q2 2026 results, CoreWeave generated approximately $2.58 billion of revenue, up 112% year over year. Revenue backlog reached $104.2 billion, up from $99.4 billion one quarter earlier, before more than $25 billion of additional customer commitments secured early in Q3. Management said near-term capacity was effectively sold out and that new agreements were being signed on increasingly favorable terms.

CoreWeave subsequently raised full-year 2026 revenue expectations to approximately $12.4 billion to $13.2 billion and increased expected capital expenditures to $35 billion to $39 billion.

Those numbers matter on their own.

But they matter more when viewed alongside the token market.

The external market data are telling us:

• Token prices ↓ Token consumption ↑↑↑

• Enterprise AI spending ↑↑

• GPU rental pricing firm

• Cloud revenue ↑↑

And CoreWeave is simultaneously telling us:

• Capacity sold out

• Pricing improving

• Backlog expanding

• CapEx increasing to meet contracted demand

That is a meaningful convergence of signals.

Read more

In the late 1970s, Wall Street’s accepted wisdom was that companies below investment grade did not belong in the public bond market. The speculative-grade bonds that existed were mostly fallen angels: companies that had once been investment grade and were later downgraded.

Michael Milken challenged that convention. His insight was not that weak companies were secretly safe. It was that risk could be financed at the right price. If a portfolio of below-investment-grade bonds offered enough additional yield to absorb expected defaults and recoveries, the category could be investable even when individual credits failed.

That idea, and the original-issue high-yield market Drexel Burnham Lambert built around it, changed corporate finance. As chronicled in The Predators’ Ball, high-yield bonds gave smaller companies, leveraged-buyout firms and corporate raiders access to capital once reserved for the largest investment-grade borrowers. Credit stopped being merely a defensive allocation. It became an instrument of corporate control.

The excesses eventually became inseparable from the innovation. Drexel collapsed. Milken went to prison. The junk-bond market seized. But high yield did not disappear, because the underlying economic function was real. The market went through a credit cycle, weaker structures were exposed, capital changed hands, and a durable asset class emerged from the wreckage.

That history matters now.

Private credit is experiencing its first real test as a mass-market asset class. The headlines are ugly, liquidity is being questioned and problem loans are still migrating into non-accrual. Yet the more important question is not whether additional credit losses are coming. They are. The question is whether today’s stress represents the death of the asset class or the painful transition into a better market.

I believe the bottom is forming.

Read more

A few months ago, I closed out LRMI for the summer with a simple set of standing orders:

Negative AI headlines will come this summer. Every single one is a mispricing event.

We certainly got the volatility. Which creates excellent opportunity.
AI spending fears, capex concerns, crowded positioning, and plenty of headlines questioning whether the trade had finally gone too far. For stretches, the market gave investors every reason to doubt the theme.

And then, this past week, AI came roaring back.

The larger point from that final note still stands. Volatility does not automatically mean the underlying investment thesis is broken. The infrastructure buildout continues, capital continues to move into the ecosystem, and the opportunity set across AI remains enormous.

After spending the summer away from LRMI, I am officially back.

I have a long list of ideas I have been waiting to dig into, and plenty of new ones that came from what I learned over the past few months. Expect LRMI to get moving again very soon.

The floor is open.

If there are any companies, themes, markets, or ideas you would like me to cover, leave them in the comments below. I would love to hear what you want to see next.

Subscribe now

Disclaimer: This is for educational purposes only. Not financial advice.

I’m writing this from my gate at the airport. Bag checked, boarding pass pulled up, coffee in hand. Today, I deploy.

Starting this week, I’ll be joining a Sales & Trading desk at a top Wall Street bank as a summer intern. For those unfamiliar with the role: this isn’t a research library or a PowerPoint factory. I will be on a live trading floor: real markets, real risk, real decisions, in real time. The rotations will span within Fixed Income. I would share which desk’s but it’s [REDACTED]. It is, without exaggeration, a once-in-a-lifetime opportunity to learn what markets look like from the inside.

I am calling it what it is: I am now a Forward-Deployed Trader.

A note on the hiatus. Out of respect for compliance and the integrity of the role, I will be going dark on the Substack for the duration of the 10-week internship. No articles, no deep-dives, no macro takes. Until I receive clarity from compliance and the internship concludes. The desk comes first.

I know what this community means. The growth over the last few weeks has been nothing short of incredible. The engagement, the DMs, the feedback on the research. We are building something real here, and this summer is a chapter, not an ending.

Subscription pause → important: Starting July 1st, all active paid subscriptions will be paused and will not resume billing until the internship concludes. You will not be charged during this period. We will pick back up right where we left off.

Thank you. Genuinely. The support this community has shown has been insane, and it fuels everything I’m building with LRMI. We will return sharper, better-connected, and with a perspective that only a live trading floor can provide.

Now, a few final words before I go dark.

Standing orders for the summer

We are not in a bubble. Do not let fear-mongering convince you otherwise. Negative AI headlines will come this summer — every single one is a mispricing event. We are living through one of the greatest investment cycles of a generation. Maintain exposure across semiconductors, data centers, and the full AI supply chain, and you will be properly allocated.

The data center buildout is infrastructure, not speculation. The capex is committed, the demand signal has not broken, and the supply chain names are real. There will be weeks where a headline causes a 5% drawdown in your semiconductor basket. That is not a warning… that is an opportunity. Don’t sell the future because a journalist had a slow news day.

Until the floor closes for summer, stay long, stay patient.

quantLR, going dark.

Subscribe now

Disclaimer: This is for educational purposes only. Not financial advice.

• Bernstein 42nd Annual Strategic Decisions Conference: & both spoke at.

• TD Cowen’s 54th Annual TMT Conference: .

• Jefferies Software, Internet & AI Conference:

These conferences offer valuable color on how businesses are performing, management’s forward-looking outlook, and a strong pulse on what’s happening across the industry through the lens of company leadership.

In today’s post, we’ll dive into commentary from this week’s conferences, highlight some of the most important takeaways, and explore a few pieces of alpha the broader market may be overlooking. Enjoy!

Read more

So we tried something different. I fed six quarters of earnings transcripts, investor days, and analyst Q&A sessions across eight companies into an AI-assisted forensic framework: Primary memory makers plus the testing, chemistry, and equipment suppliers in the stack beneath them.
The Goal: triangulate what the Big 3 (Micron, Hynix, & Samsung) won’t say directly by reading the behavior of the companies paid to process their failures.

Here’s what we found.

Read more

One of my favorite ways of conducting research is through what I call Forward-Deployed Intelligence, getting offline, going outside, and talking to people who are actually inside the machine.

Mosaic investing is one of the greatest edges available in modern markets. Not because the information is secret, but because most people never bother to go get it.

This week I caught up with a former coworker over beers. He’s a Memory OCM, sitting inside the procurement layer of one of the largest tech supply chains on earth, and what he shared painted a picture that I don’t think the market has fully realized.

It’s a jungle out there! I’ll be your jungle guide today.

Before we dive in… the supply chain color in this piece comes from a conversation with a former colleague working as a Memory OCM: a procurement-layer role with visibility into lead times, allocation dynamics, and pricing sentiment — not corporate financials or forward guidance. This is mosaic research, the same method institutional investors pay expert networks thousands of dollars per hour to access. For educational purposes only. Full disclosure at the end.

Culture is King, Trust is Second to None

Memory supplier stocks have absolutely skyrocketed — Micron +683% over the past year, SK Hynix +870%, Samsung +439%.

Read more

Four of the six metric families are re-accelerating simultaneously at a scale that defies the law of large numbers. The two that appear to be "decelerating" — gross margin and operating margin improvement rates — are decelerating into all-time record levels, not into deterioration. This is not a normal earnings pattern. This is a demand supercycle with compounding operating leverage, and the second derivative across revenue, data center, EPS, and FCF has not yet peaked!

LRMI Financial Physics

A quick note for new readers: What is Financial Physics?

In physics, the first derivative of position is velocity — how fast something is moving. The second derivative is acceleration — how fast the velocity itself is changing. Most Wall Street analysis lives entirely in the first derivative. Revenue grew 80% — good. EPS beat by 5% — great. Stock goes up.

I personally believe that’s not enough.

What separates elite macro and fundamental investors, Druckenmiller, Soros, Simons from the consensus is that they think in second derivatives. They don’t just ask “is this good?” They ask “is it getting better or worse, and at what rate?” A company growing at 80% YoY but decelerating from 120% is a fundamentally different animal than a company growing at 80% and re-accelerating from 50%. The price action will eventually reflect that difference… but the market is almost always late.

I call this framework Financial Physics because it borrows directly from classical mechanics: position, velocity, acceleration. Applied to earnings, it looks like this:

• Position → the absolute level (revenue of $44B, EPS of $X)

• Velocity → the rate of change (revenue +69% YoY)

• Acceleration → the rate of change of the rate of change (revenue growth re-accelerating from +52% → +69% → +93%)

When acceleration is positive and building across multiple metric families simultaneously → that is the signal most investors miss entirely. That’s what I look for. And that’s exactly what Nvidia just printed!

NVIDIA | 2nd derivative full breakdown — Q2 FY26 → Q1 FY27

1. Data Center revenue — ↑ Re-accelerating

The Data Center 2nd derivative story is actually stronger than the consolidated revenue story. After a +10pp acceleration in Q3 that dipped to +9pp in Q4 (Blackwell transition stutter), Q1 FY27 printed a +17pp jump to +92% YoY. That Q4 “dip” in the delta was the setup. The market read it as deceleration. It was inventory timing on the Blackwell ramp. The Q1 FY27 snap-back reveals it.

Read more

You may have heard: Cerebras ($CBRS) IPO’d this week and has gathered a lot of attention. For those unfamiliar, Cerebras is primarily known for developing the world’s largest and fastest AI computer chips. Uniquely designed to accelerate AI training and inference by using entire silicon wafers rather than cutting them into smaller, individual processors. Their unique approach aims to eliminate the data-movement bottlenecks of traditional GPU clusters, with a focus on delivering high-speed, low-latency AI performance.

SemiAnalysis is one of my favorite institutional-grade semiconductor publications in the world! Their work is consistently top-tier and gives excellent market color for the semiconductor industry and AI infrastructure.

This past week, they dropped a deep-dive: “Cerebras — Faster Tokens Please.” And while reading it, I realized something the rest of the market is going to miss:

Cerebras is long memory.

This post is tailored to walk you through how, why, and what it means for the Memory market. Enjoy!

SemiAnalysis

Cerebras — Faster Tokens Please

It’s been nearly 5 years since Dylan wrote a dedicated article about Cerebras in June of 2021 for the newsletter. He shipped 4 articles in 2 days! How times have changed…

Read more

5 months ago · 135 likes · 6 comments · Myron Xie, Jordan Nanos, Max Kan, Cam Quilici, Tanj Bennett, Ivan Chiam, Louis Lu, Zane Fong, Gerald Wong, Reyk Knuhtsen, Nicolas Bontigui, Wega Chu, Dylan Patel, and Konrad Wang

BLUF: Bottom Line Up Front

Read more

My commodity covered the full spectrum: HBM, NAND, SSD, DDR4, DDR5 (which had just hit the market), and even DDR6 (already in testing, years ahead of production). My suppliers? They’re the names you see trading today. At 19, I was getting a real Semiconductor 101 education, not from a textbook, but from the supply chain itself. (I was a Finance Major lol).

ChatGPT was going mainstream. NVIDIA was trading at $1,000 pre-split. I had a hunch, even a conviction, that semiconductors and memory would be foundational to the AI boom ahead. But I’d be lying if I said I fully grasped the magnitude of what was coming.

I could hear the music when no one else could. It was beautiful. Fast forward to Q3-Q4 2025: the rest of the world is finally hearing the melody I’d already caught. Today, I’m here to tell you that same song is getting louder. We’re very early to what I’m now calling The Memory Permanence.

Memory was cyclical for years. Now it's permanently structural.
I believe we are only in the 3-4 innings of The Memory Supercycle, this post will share my view on Memory. Enjoy!

The Number Nobody Is Talking About

This might be the only place you'll see this exact computation modeled this week. That's the edge you're paying for.

Let me give you one data point and let it sit for a second.

Read more

I’ve been skeptical of that framing for a while. And this week, Citadel Securities published a piece that puts the institutional weight of one of the most rigorous macro research shops on Wall Street squarely behind the contrarian view.

Subscribe now

The title: The Economics of Intelligence.

The conclusion: AI is a complement to labor, not a substitute — and the data are already proving it.

What Citadel Actually Found

Frank Flight and the Citadel Securities macro team did something most AI commentators haven’t bothered to do: they looked at the actual data.

Their analysis ran NLP-based keyword classification across S&P 500 earnings call transcripts, tracking how corporate executives have framed the relationship between AI and their workforces over time. The results are striking.

AI mentions on S&P 500 earnings calls grew from 20% to 64% of companies — a near-tripling that began accelerating sharply after ChatGPT launched in early 2023. If AI were actually gutting headcount, you’d expect that acceleration in AI adoption to correlate with aggressive labor cutting signals.

It doesn’t.

Labor cutting language — layoffs, severance, hiring freezes — peaked at 21% in Q1 2023 and subsided to 13% by Q4 2025. Hiring and talent language has remained steady. Restructuring language hasn’t budged. The AI adoption wave, in the actual corporate data, has not produced a labor market shock.

But here’s the number that stops me cold:

When executives mention AI alongside their workforce, they frame it as a complement to labor vs. a substitute by a ratio of roughly 8 to 1. 43% of firms call AI a complement. 51% are neutral. Only 5 → five percent, frame AI as a substitute for human labor.

These are not Silicon Valley futurists or ideologically motivated academics. These are CEOs and CFOs on earnings calls, speaking to institutional investors, where credibility is literally priced in. They are telling you, under that scrutiny, that they are using AI to make their existing people more productive — not to replace them.

The Physics Argument: Why Compute Intensity Changes Everything

Citadel’s framework builds on a structural argument that I find compelling from first principles, and one that maps directly onto my own research from earlier this year on AI infrastructure constraints.

Subscribe now

The core thesis has four linked components:

1. Compute demands scale quadratically with task complexity. This is not intuitive, but it’s crucial. The jump from automating a simple task — summarizing an email, generating a code snippet — to automating a complex, multi-step cognitive role is not linear. It requires orders of magnitude more compute. The physical constraints are real: silicon, power, cooling, data center buildout. These don’t scale overnight.

2. Physical scarcity limits generalized displacement. Because complex cognitive work is expensive to automate at scale, the economics favor augmentation over wholesale replacement. It is cheaper and faster to give a skilled worker an AI co-pilot than to build and run the compute stack that could fully replace them.

3. AI therefore integrates into existing workflows as a force multiplier. This is the radiologist story. This is the lawyer who drafts in half the time. This is the analyst who can synthesize ten earnings calls before the market opens. The tool makes the human more productive without eliminating the human’s indispensable judgment, relationships, and accountability.

4. History says this is the pattern — and demand is elastic. Technological revolutions don’t eliminate labor as an input. They alter the task composition of jobs while output expands to fill the new productive capacity. Microsoft Office didn’t replace office workers. It redefined what office workers could do — and the economy created more of them.

As Citadel puts it directly: “Was the advent of Microsoft Office a complement or substitute for office workers? Ex-ante the concern skewed towards substitution, ex-post it appears a clear complement.”

That’s the bet we’re making again, and the early data are on the right side of it.

Jevons Paradox: The Demand Expansion Nobody Is Modeling

There’s a concept embedded in Citadel’s analysis that I think is the most underappreciated macro dynamic in the AI debate: Jevons Paradox.

Named after 19th century economist William Stanley Jevons, who observed that more efficient coal-burning engines actually increased total coal consumption — because efficiency drove adoption and expanded use — the paradox describes how productivity improvements in a resource often increase total demand for that resource rather than reducing it.

Applied to AI and labor: as AI makes workers faster and more productive, the economy doesn’t simply need fewer workers. It expands the scope of what gets done. More tests get run. More clients get served. More software gets built. The demand for skilled human judgment scales with the expansion of output, not against it.

We’re already seeing this in the construction labor market. AI data center buildout is driving a measurable pickup in construction hiring. The infrastructure layer of the AI revolution is physically employing people. This is Jevons working in real time.

Jensen Has Been Saying This for Months — And He Was Right

I want to give credit where it’s due. Jensen Huang has been in front of this thesis longer than anyone in the financial media has taken seriously.

His radiologist example has become something of a signature argument, and for good reason — it is devastating to the displacement narrative.

Ten years ago, Geoffrey Hinton — the so-called “Godfather of AI” — predicted that AI would make radiologists obsolete. His technical forecast was correct: AI is now embedded in nearly every corner of radiology. But the labor outcome was exactly the opposite of what he predicted.

As Huang explained at the Milken Institute this week: “People who believe this misunderstand that the purpose of a job and the task of a job are related but not ultimately the same thing.”

The radiologist’s job is to diagnose disease. Reading scans is a task in service of that purpose. When AI made image analysis faster and more precise, radiologists could read more scans, serve more patients, and generate better economics for hospitals. The response? Hospitals hired more radiologists. The American College of Radiology projects the number of radiologists in America will grow by up to 40% between 2023 and 2055.

Huang has extended this logic broadly. In his recent appearance at the Milken Institute: “AI creates jobs. AI is the United States’ best opportunity to re-industrialize itself.”

And the data is starting to confirm it. Software engineer job postings on Indeed were up 11% year-over-year in early 2026 — in the middle of what was supposed to be an AI-driven tech layoff wave. Huang himself noted that AI has created more than half a million jobs in recent years.

“I’m fairly confident that AI will drive productivity, revenue growth, and therefore more hiring,” he said.

The inflection is happening. We’re watching it in real time.

The Macro Positioning Read

For those of us managing capital, what does this framework actually imply?

Citadel’s analysis makes an argument that I think has real alpha embedded in it: the market has been pricing AI as a labor destroyer, and that consensus is wrong. Sectors like Software have been brutally sold-off the past year; which if you believe in this color — then software presents an amazing opportunity. As evidence accumulates that AI adoption is not producing the employment shock the doomers forecast, several second-order effects follow:

• Wage growth remains stickier than feared → inflationary impulse doesn’t disappear the way bears assumed

• Consumer spending holds up better than recession models embedded in AI-displacement scenarios → consumer staples and discretionary don’t get the demand shock

• AI infrastructure investment remains front-loaded and massive → the real challenge is the payoff structure (decades of productivity gains vs. near-term capex) not the demand destruction narrative

• The big winner remains the infrastructure layer: compute, power, data centers — the physical real estate of intelligence

The Citadel piece explicitly flags what I think is the real risk to the AI trade: not that AI destroys jobs, but that the investment cycle is heavily front-loaded while the productivity payoff accrues over decades. That’s a very different bear case than the labor displacement narrative — and it’s a bull case for the physical infrastructure names I’ve been long. & .

The pie is getting bigger. The question for investors is simply: who has the biggest fork?

Disclaimer: This is for Educational Purposes Only. NFA.

The discount was rational. Korean corporate law genuinely did not protect minority shareholders. Directors owed fiduciary duty to “the company,” which in practice meant the controlling family. Treasury stock was used as a financial engineering tool to entrench control. Mergers happened at unfair ratios. Cross-shareholdings hid value. The market priced this correctly. The Korea Discount was not a mispricing, it was an accurate read on the legal architecture.

What changed in late 2025-2026 is that the legal architecture changed.

The regime change

On July 22, 2025, an amendment to Article 382-3 of the Korean Commercial Code took effect. It explicitly extends director duty of loyalty to shareholders, not solely to the company. The vote was 220 to 272 — not symbolic, statutory. On August 25, 2025, a second amendment expanded the aggregate 3% rule for audit committee elections, renamed outside directors as “independent directors,” and mandated hybrid AGMs from January 1, 2027. The FSCMA enforcement decree, effective December 31, 2024 — prohibited treasury stock allocation in mergers and spinoffs, closing the most abused governance loophole in modern Korean capital markets.

The Corporate Value Up program, launched February 2024, has now crossed 590 disclosure filings. KRX announced on April 15, 2026 that the simplified disclosure path will be eliminated in 2027 — full plans with status diagnosis, target setting, implementation timeline, and communication strategy will be required. KRX will publish a “low-PBR list” in H2 2026 designed to name and shame companies that have not engaged.

Korean activists now hold real war chests. Align Partners runs ₩1.235 trillion in AUM and has filed AGM proposals at six small-caps in the last 12 months. Dalton Investments has accumulated 5%+ stakes across multiple cosmetics, holdco, and mature-industrial names. Truston has campaigned at BYC since 2021. KCGI executed the first ownership change at Hanyang Securities in 68 years.

These are not promises. They are enacted laws, court precedents, and active capital deployments. The Korea Discount existed because the legal architecture justified it. The legal architecture is now changing faster than institutional pricing has adjusted.

That gap is the trade.

What I’m doing with personal capital

I’m currently building a 15-name basket of Korean equities across three theses: AI infrastructure suppliers selling into the NVIDIA, Micron, SK Hynix, Samsung and TSMC supply chains; deep-value small caps trading below tangible book with named institutional activists on the register; and founder-succession setups where inheritance tax of 50-60% on controlling shares structurally forces value-realizing events.

I’ll be deploying capital into these positions over the coming weeks, building the most liquid names first and accumulating the smaller positions over 4-8 weeks. This is personal capital with a 5-10 year horizon, not a short term trade. Position sizes and entry timing will be a function of liquidity constraints, not narrative urgency.

I’ll be transparent about which names I purchase and at what levels. Not all 15 will be sized equally. Some I’ll own forever. Some I’ll own for 24 months around a specific catalyst. The basket is calibrated to a regime change, not a single thesis.

Tier 1: AI infrastructure (the picks-and-shovels under NVIDIA and TSMC)

Read more

These seven charts, read together, paint a coherent and unsettling macro picture. This isn’t a soft patch — this is a regime transition with compounding signals across correlations, supply chains, capex, and FX. Here’s the breakdown:

Subscribe now

1. AI Dispersion (Chart 1) — The AI Trade Has Bifurcated

The 6-month pairwise correlation across /// has collapsed to ~0.20–0.35, the lowest since at least 2013. This is one of the most important structural signals in the chart deck. The “AI basket” trade is dead as a monolith. The market is now doing fundamental differentiation — pricing MSFT’s Azure growth differently from GOOGL’s search disruption risk, etc.

Implication: You can no longer be long “AI exposure” as a factor. You have to be right on the specific company. This also means dispersion vol is the trade — long individual names via options rather than basket hedges. The single-stock picker’s market is back.

2. Allied Defense Spending (Chart 2) — The Rearmament Supercycle Is Accelerating

Germany and Japan are the standout stories — near-vertical ramps into 2025–2026. Germany crossed ~$100B (nearly doubling in 3 years), Japan accelerating rapidly off a low base. This is a decade-long structural spend shift, not a cyclical bump.

Implication: Defense primes (, , , BAES.L, Rheinmetall) and European defense ETFs (NATO proxy plays) remain multi-year longs. The market has partially priced this but the duration of the spend is underappreciated. This is 10-year capex commitments, not 2-year. My personal favorite way to play the European Rearmament is (Euro Defense ETF).

3. Europe Jet Fuel Supply Shift (Chart 3) — Energy Supply Chain Rewiring in Real Time

European jet fuel/kerosene imports show Middle East supply cratering in April (barely ~100K bbl/day vs. 300K+ earlier) with North America filling the gap at ~200K bbl/day. This is almost certainly Iran sanctions/Hormuz disruption driven. Another win for the US and the Trump Admin.

Implication: The Hormuz choke-point risk is real and already showing in physical flows. This is not priced in airline operating costs, and it’s bullish for US refined product exporters , and LNG shippers. European energy security premium is rising.

4. Core Capital Goods Orders (Chart 4) — Capex Surge, But Read It Right

March 2026 core capital goods orders (ex-aircraft, ex-military) surged ~3.5% MoM — the biggest print since COVID. This looks bullish on the surface, but context matters: this likely reflects front-loading ahead of tariffs, not genuine organic demand expansion.

Implication: Don’t read this as a capex cycle inflection. It’s a pull-forward. The second derivative will likely turn negative in Q2–Q3 as the front-loading exhausts and tariff uncertainty paralyzes forward commitments. I’d watch the next 2–3 prints closely. Industrial machinery and equipment names could give back recent gains.

5. Stock-Bond Correlation (Chart 5) — The Most Dangerous Chart in the Deck

The S&P 500 / Treasury return correlation has flipped positive and is now running near +0.15 to +0.20 — a regime that has historically only persisted during inflationary supply shocks (1970s, 2022). The traditional 60/40 hedge is broken. Bonds don’t protect when stocks sell off in an inflation regime.

Implication: This validates my personal “Broken 60/40” thesis and it means the Fed is pinned — cutting into an inflation scare kills the bond rally and doesn’t help equities. Gold, commodities, real assets, and short duration are the only clean hedges. This chart alone justifies a significant portfolio construction shift.

6. Housing Starts (Chart 6) — Positive Surprise, But Tariff Shadow Looms

Total housing starts hit 1.5M SAAR — the best in over a year, with single-family at ~1.0M. This is a genuine beat and signals that builders are getting ahead of expected tariff-driven lumber/materials cost increases.

Implication: Same dynamic as capex orders — likely a pull-forward, not a trend. Lumber futures and homebuilder stocks , , may have a near-term ceiling as this front-loading exhausts. The real question is mortgage rate trajectory — if the stock/bond positive correlation persists, rate relief is not coming, which chokes off sustained housing demand.

7. USD/JPY and BoJ Intervention (Chart 7) — Japan at 160, No Intervention Yet

Dollar-yen is back near 160 in 2026 — the same level that triggered massive BoJ intervention in 2024 (¥5.62T and ¥5.92T interventions). Critically, no intervention has occurred yet at this level despite the yen sitting right at the prior trigger zone.

Implication: Either (a) the BoJ has decided to tolerate more weakness given US tariff/trade dynamics, or (b) intervention is imminent and the risk/reward on USD/JPY is capped at current levels. For a US macro book, this matters because yen weakness = dollar strength, which is a headwind for EM and commodities priced in USD. We should all watch for a coordinated G7 FX statement — that would be the catalyst for a sharp yen snap-back.

Synthesis: The Big Picture

1. AI Dispersion Collapse — Late-cycle differentiation, monolith trade over

2. Defense Ramp — Multi-year structural, not cycle-dependent

3. Jet Fuel Rerouting — Hormuz risk real, energy supply chain fracturing

4. Capex Surge — Front-loading, not organic — second derivative turns

5. Stock/Bond Correlation Positive — Inflationary regime, 60/40 broken

6. Housing Beat — Pull-forward, rate-dependent sustainability

7. USD/JPY Near 160, No Intervention — Intervention risk building, dollar cap forming

Bottom line: It appears we are in a late-cycle inflationary supply shock regime with geopolitical overlays compressing traditional hedges. The plays are: long real assets/energy/defense, short the 60/40 assumption, be selective (not basket) in AI, and watch USD/JPY as the canary for FX intervention risk.

Disclaimer: This is for educational purposes only. NFA.

Two charts came across my feed this morning from Apollo Chief Economist Torsten Slok — and together they tell one of the cleaner second-derivative macro stories I’ve seen in a while. The market hasn’t caught up yet.

Subscribe now

Chart one: median wage growth is bending back up. After cooling from the post-COVID peak, the job-switcher series — historically the leading indicator for broader wage moves — is re-accelerating into early 2026. Chart two: net immigration has collapsed from 2.8 million in 2024 to a CBO-projected 0.4 million in 2025, recovering only modestly to 0.6–0.7 million through 2027.

Put them together and the story is simple: a labor supply shock is tightening the market in exactly the sectors that relied most on immigrant workers. Native-born workers in construction, agriculture, meatpacking, and hospitality are starting to see real wage gains. And those workers vote.

“The market is pricing the generic ballot. The pocketbook voter is pricing their paycheck.”

The Political Transmission Mechanism

Pocketbook voting is one of the most robust findings in political science. When real wages rise visibly, the incumbent party benefits. Reagan ‘84, Clinton ‘96, the 2019 low-income wage surge all confirm it. The question isn’t whether the mechanism exists — it’s whether the wage acceleration is large enough, fast enough, and felt by the right voters to shift the outcome in swing districts.

The Slok data suggests the answer is yes on all three. The job-switcher series is your early signal — it moves first, stayers follow with a 3–6 month lag. We are currently in that lag window. By September and October, when early voting starts, the wage improvement will be visible in paychecks across the states that matter most.

The critical variable the prediction markets are missing: this isn’t a national story. It’s a state-specific labor market story. And some states are dramatically more exposed than others.

The State-by-State Mispricing Map

If you're interested in trading any of these markets, consider using my referral link to sign up for Polymarket → it helps support independent research like this. 🔗

https://polymarket.com/?r=Kairos7

Texas Senate

This is the most egregious single misprice on the board. Texas has one of the highest concentrations of immigrant labor in construction, oil field services, and agriculture in the country. An 86% collapse in net immigration hits Texas working-class voters hardest and most visibly — and the wage uplift in those sectors will be felt first.

Democrats are currently priced at 43 cents to win the Texas Senate race. Democrats have not won a Texas statewide race since 1994. Thirty years. I have this as 25–28 cents fair value at most. That’s a 15–18 cent mispricing on a $183K volume market — liquid enough to size a real position.

Iowa Senate

Iowa is ground zero for this thesis. Meatpacking and agricultural processing dominate the Iowa interior — Tyson, JBS, and similar facilities employed enormous immigrant workforces. With immigration collapsed, native-born wage pressure in those plants is real, immediate, and politically legible to the voters who work there.

GOP is currently priced at 60 cents. I have fair value at 70–72 cents. The market is underpricing rural Iowa’s sensitivity to exactly the labor dynamic the Slok charts describe. Buy the Republican Iowa Senate position.

Alaska Senate — Mary Peltola vs. Dan Sullivan

This is the cleanest individual race mispricing on the board. Peltola is priced at 64 cents — Sullivan at 36 cents — in a state Trump won by 13+ points in 2024. The market is giving significant RCV incumbency premium to Peltola, but Alaska’s fishing and resource extraction industries are not insulated from the labor dynamics in play.

More importantly: Alaska at 64% Democrat in the current political environment is simply wrong on priors alone. I have Sullivan at 45–50 cents fair value. That’s a 9–14 cent edge on the highest-volume individual Senate race on the page ($306K).

Ohio Senate

Ohio is the interesting swing case. Manufacturing and construction workforce with meaningful immigrant labor overlap, in a state Trump won by 11 points in 2024. Democrats are priced at 58 cents here — a slight favorite. If the wage acceleration story plays out in the Ohio Rust Belt, this flips. Medium conviction, 8–10 cent edge toward GOP.

The Full Alpha Stack

What Kills This Trade

Two things break the thesis and you should watch both closely.

First: tariff-driven goods inflation. The April 2026 tariff regime on steel, autos, and broad consumer goods could re-ignite inflation and compress real wage gains back toward zero. Nominal wages up 4% plus goods prices up 3% equals a real gain of roughly 1% — which voters don’t feel. Watch the April CPI print (May release) as a pivot point. If goods prices reaccelerate, the Slok thesis weakens as a political argument.

Second: sector-specific unemployment. The immigration collapse has two sides. Tighter labor supply pushes wages up, but it also slows hiring in construction and agriculture. If regional unemployment spikes in immigrant-dependent sectors — Texas homebuilding stalls, Florida agriculture contracts — the job loss narrative will dominate the wage gain narrative in local media and voter sentiment.

The Data to Watch:

Atlanta Fed Wage Growth Tracker (monthly), BLS JOLTS job openings in construction + leisure/hospitality, generic ballot by state (not national), and April CPI goods ex-food/energy. When the generic ballot in Texas or Iowa tightens toward Dem +2 or better, that's the signal the market is repricing — and your position is moving.

The prediction markets are running a national model on what is fundamentally a disaggregated labor story. The crowd is pricing the generic ballot from three months ago. The edge is in being early on the second derivative — the rate of change in real wages for the workers in the states where the immigration-to-wage transmission is fastest.

The market will be the last to know. That’s where the alpha is.

Intel delivered its sixth consecutive quarter beating expectations, with Q1 2026 revenue of $13.6 billion, up 7% YoY.

The beat was broad-based — revenue came in$1.4 billion above the midpoint of guidance, driven by supply improvements, better mix, and pricing actions.

Subscribe now

The GAAP loss was distorted by a $3.9 billion non-cash goodwill impairment charge on Mobileye, driven by increased discount rates and macro uncertainty — not an operational deterioration. AI-driven businesses now represent 60% of revenue and grew 40% YoY!

Q2 2026 guidance: Revenue of $13.8–$14.8 billion, non-GAAP gross margin of 39.0%, non-GAAP EPS of $0.20. DCAI guided to grow double digits sequentially.

LRMI: Analysis: Data Center & AI Infrastructure Demand

I believe this is where the real signal is. Management was unusually direct about a structural shift in the CPU-to-accelerator compute ratio, and the data center numbers confirm it.

DCAI Segment — The Core Thesis

• DCAI revenue: $5.1 billion, up 22% YoY and 7% sequentially, described as “well above expectations.”

• ASIC revenue nearly doubled YoY and was up >30% sequentially — an underappreciated growth vector.

• Operating profit: $1.5 billion at 31% margin, up ~$292M QoQ on better yields and cycle times on Intel 3.

The CPU-to-GPU Ratio Inflection — The Most Important Data Point on the Call

Management quantified what the market has debated: the ratio of CPUs to GPUs in AI workloads is structurally shifting toward CPUs as AI matures:

• Training: ~7–8 GPUs per 1 CPU

• Inference: now closer to 3–4:1

• Agentic / multi-agent: approaching parity or better for CPUs

CEO Lip-Bu Tan reinforced this: “The backbone of AI computing in production remains a CPU-anchored architecture... CPU is the orchestration layer and critical control plane for the entire AI stack.”

Inference requires CPU-heavy orchestration for managing multi-agent data flows where power-per-performance matters.

Demand Signals Are Getting Stronger, Not Weaker

• Server CPU unit TAM outlook improved over the last 90 days — Intel now expects double-digit unit growth for the industry and for itself, with momentum extending into 2027.

• Supply, not demand, is the binding constraint. Management explicitly stated demand continues to outpace their growing supply.

• To address this, tool spending is increasing ~25% YoY even as total CapEx stays flat — a signal of deliberate capacity acceleration on the manufacturing side.

Customer Wins — Structural, Not Transactional

• Google signed a multi-year long-term supply agreement for Xeon IPU in Q1, with Lip-Bu describing it as a “long-term trusted partnership.” More contracts to be announced.

• NVIDIA’s DGX Rubin NVL8 selected Xeon 6 as the host CPU — Intel is embedded in the most advanced AI training systems being shipped today.

• SambaNova signed a multiyear collaboration for next-gen heterogeneous AI inference architectures combining SambaNova RDUs with Xeon 6.

Advanced Packaging — A Scarce and Strategic Asset

The semiconductor TAM is approaching $1 trillion driven by AI demand. Intel has identified three strategic moats: the x86 CPU franchise, advanced packaging, and the manufacturing network.

Advanced wafer and packaging capacity remains inshort supply industry-wide, which benefits Intel’s vertically integrated supply chain. The company is expanding its back-end facilities in Penang, Malaysia on a multi-year horizon.

Near-Term Gross Margin Watch

CFO Zinsner flagged that Q2 gross margin will step down from 41% to 39% due to Panther Lake (18A) volume ramping ~6–7x QoQ — still below corporate average margins. Input cost inflation (substrates, T-glass, memory) is also a headwind in H2. This is a ramp tax, not a demand problem.

Stanley Druckenmiller Variant Perception

Using Druckenmiller’s framework: find the consensus that is wrong, identify the leading indicator the market is missing, and size the bet on the rate of change, not the level.

Subscribe now

What the Market Has Wrong

The market’s dominant narrative has been “GPU wins AI, CPU is commoditized legacy silicon.” Intel has traded as a restructuring story, not a growth story. The variant perception emerging from this call is that this framing is 3 years stale.

The Asymmetric Setup

1. The CPU TAM is being re-rated in real time. The market modeled GPU/accelerator-centric AI infrastructure. The shift to inference and agentic AI is a second derivative change that dramatically increases CPU intensity per unit of compute output. Intel’s own numbers — DCAI +22% YoY, server CPU TAM now expected up double-digits vs. flat expectations 90 days ago — are the leading indicator. The rate of change in estimates is what Druckenmiller would focus on.

2. Supply is the constraint, not demand. When demand exceeds supply and the company is accelerating tool spend 25% YoY, you have a pricing power and revenue upside setup into H2 and 2027. The market is discounting Intel on margin concerns (18A ramp costs) without properly crediting the demand backlog building underneath.

3. 18A is a free call option the market is not pricing. Yields are running ahead of internal projections — described as a “meaningful inflection.” If 18A yields continue to improve and external foundry customers commit (Intel has flagged that 14A continuation depends on external design wins), Intel goes from a one-asset story (x86 CPUs) to a two-asset story (x86 + foundry), re-rating the stock.

   

4. The ASIC business is hiding in plain sight. ASIC revenue nearly doubled YoY and grew 30%+ sequentially. This is structurally attached to the custom silicon wave (hyperscalers building their own chips). Intel’s packaging and manufacturing capabilities make it a natural beneficiary. The street is not modeling this segment discretely.

5. Consensus is modeling PC weakness as a drag. Full-year PC TAM is expected down low double digits and H2 client is weak. This is the consensus anchor weighing on Intel’s multiple. But CCG is now >60% AI PC mix. As AI PC penetration grows and the platform refreshes, the downside case on CCG may be more muted than feared.

The Risk That Kills the Variant View

The bear case is straightforward: 18A fails to attract sufficient external design wins, making the foundry investment a value destroyer with no revenue to justify it. Management explicitly flagged that 14A development could be paused if external customer demand isn’t secured.

Additionally, input cost inflation (substrates, memory, helium shortages) compressing margins in H2 could cause estimate cuts that overshadow the demand recovery narrative. The bet requiresprocess execution— which Intel has historically struggled to deliver on time.

Bottom line: The earnings call revealed a company where the demand picture is materially better than expected and improving, the CPU structural narrative is gaining real-world confirmation, and the primary risk is self-inflicted execution (18A ramp, costs) rather than market demand. The variant perception is that Intel is being priced as a restructuring story when the data now supports framing it as an AI infrastructure beneficiary in an early innings re-rating.

With CPU plays like , , and surging +40% in just two weeks, it’s clear the market is aggressively correcting a previous mispricing in the space. I anticipate a rotation of profits back into GPU leaders like , whose May 20 earnings (T-24 days) will serve as a definitive litmus test for both GPU demand and the broader AI sector.

Disclaimer: This is for Educational purposes only. Not a solicitation to buy or sell securities. NFA.

• Forward Observations Desk is a field-deployed intelligence function within an investment organization that gathers real-time, on-the-ground insights → through site visits, and supply-chain checks— to inform trading decisions and validate or challenge market narratives ahead of broader consensus.

Today’s piece is brought to you by LRMI’s Forward Observations Desk, Enjoy!

A SemiAnalysis tweet hit X yesterday with two GB300 dev racks CoreWeave scrambled to deploy after their cluster went down. Standard AI infrastructure stuff — until you actually look at the photo.

Read more

Revenue: ₩52.6T +198% YoY · beat est +7%

Op. Margin: 72% record · +30pp YoY

EBITDA: ₩41.3T 79% margin · beat est

Net Cash: ₩35T flipped from net debt

This quarter was not supposed to be possible

To put 72% operating margins in context: Apple, the most profitable consumer hardware company in history, runs at roughly 30%. SK Hynix — a Korean memory manufacturer that analysts have modeled as a commodity cyclical for two decades — just more than doubled that, in Q1, which is historically their softest seasonal quarter.

Revenue hit ₩52.58T, up 60% from the prior quarter and 198% year-over-year. DRAM average selling price surged mid-60% quarter-on-quarter with flat bit growth — meaning the entire revenue jump was pure pricing power, not volume. NAND ASPs were up mid-70% QoQ. This is not a demand-pull inventory cycle. This is structural scarcity pricing.

Four consecutive record operating profit quarters — Q2 '25 through Q1 '26 — with COGS rising only 6% QoQ on 60% revenue growth. That is what operating leverage looks like when a commodity business stops being a commodity.

“Customers’ demand for the next 3 years far exceeds our current supply capacity. Within the limited supply capacity, we are doing our best to supply as much HBM as possible to our customers.”

— Kim Gi-Tae, Head of HBM Sales · SK Hynix Q1 2026 Earnings Call

A three-year demand visibility statement. In memory semiconductors. That has never happened before in this industry's history.

LRMI Derivative Analysis: Where we are in the cycle: momentum, acceleration, jerk

Most earnings recaps will tell you revenue went up. Here's the more interesting question, I’ve developed: is the rate of improvement getting faster or slower? Running the three mathematical derivatives on quarterly revenue gives a precise cycle positioning signal most analysts never compute.

Read more