Is it Really?
The bubble thesis rests on three assertions: that compute is being overbuilt, that the capital behind it will not earn a return, and that the intelligence it produces is commoditising. Each is testable. We have tested all three against contract pricing, renewal economics and disclosed payback periods and we find no supporting evidence in the data. What the data shows instead is a market clearing at record prices, with capital recovery measured in months rather than cycles.
1. Rental rates are not behaving like a glut
Prices for last-generation silicon are at or near record levels, and have held there for months.
If compute were being overbuilt, the first place it would show is the rental market. It has not. On our reading of the daily rental indices, rates across the A100, H100, H200 and B200 are up roughly 30% year to date and have sustained those levels rather than spiking and fading. Silicon Data's B200 benchmark rose 24.4% in the first quarter alone, and the H200 neocloud rate rose 14.4% between early May and late July 2026 while the H100 rose 7.0%. The H100, note, being a chip two architectures from the frontier.
The distinction that matters is between tiers. Hyperscaler H100 pricing has been close to flat, anchored to long-dated capacity commitments, while the neocloud tier, where marginal demand actually clears has moved double digits. Bears quoting flat hyperscaler list prices as evidence of softening demand are reading the least informative series in the market.
The market is also institutionalising. CME will launch futures on Nvidia GPU rental rates on 5 October 2026, which turns compute into a hedgeable asset class and, for our purposes, produces a forward curve — the first observable, market-priced view of whether the glut everyone is forecasting actually arrives.
2. What the IaaS providers are actually saying
Three independent operators, the same disclosure: old chips are re-contracting at premium prices.
This is the single most under-read set of datapoints in the sector, because it goes directly at the depreciation bear case. If GPUs were stranding at four years, renewals would clear at discounts. They are clearing at premiums.
Source: company earnings calls and filings, Q2/Q3 2026. LC Equity Research.
Oracle framed it plainly on the call: capacity coming up for renewal is achieving higher prices, on the order of 20%, which management reads as a positive signal for demand, growth and the profitability of the business. CoreWeave's renewal sits against a $104 billion revenue backlog and more than $25 billion of further commitments booked since July. Contracted power of roughly 4.2 gigawatts against an installed base of 1.5 gigawatts means customer commitments already cover close to triple what the company can currently deliver.
These are three operators with different cost bases, customers and capital structures, disclosing the same thing in the same quarter. That is not management spin; that is a price signal.
A 2029 contract on 2020 silicon does not settle the depreciation debate in the abstract. It does settle whether the assets strand.
3. "Massive capital misallocation with no returns"
This is the bears' favourite claim. We would like to see the evidence, because the payback arithmetic points the other way.
Start with what capacity is contracting for. Nebius signed four AI cloud contracts in Q2, each averaging more than $1 billion in total value, at annual contract values above $20 million per megawatt — against roughly $12 million per megawatt across its existing 2026 capacity. Short-dated capacity, deliverable in three to six months, is clearing at $40–50 million per megawatt, roughly double standard one-to-three-year terms at $20–25 million. At the top end, Anthropic contracted roughly 325,000 GPUs across SpaceX's Colossus campuses for $1.25 billion a month, and Google took about 110,000 GPUs for $920 million a month. On disclosed capacity of more than 300 megawatts for the Anthropic agreement, which puts that deal at approximately $50 million per megawatt per year.
Now put that against build cost. Our working number for a greenfield gigawatt is $35–40 billion all-in; external estimates run $40–50 billion in servers, facilities, network and energy. The resulting payback periods are not close to the bear case.
LC Equity Research. Payback shown on gross capital outlay per gigawatt, revenue basis, before financing and operating costs. Illustrative.
So the range across the contracting spectrum runs from roughly ten months to a little over two years and that is at the newest neoclouds, several of which do not have a competitive software layer. Nebius put the expected payback on Q2 deals at 1 year 10 months, down from a historical two-to-three-year range.
The picture improves materially if you assess payback on the operator's own capital rather than gross project cost, which is the economically correct frame given how these projects are financed. Around 70% of Nebius's Q2 contracts carried upfront payments, with more than $9 billion of customer prepayments expected during 2026, covering 50–60% of related capital expenditure. Oracle closed more than $30 billion of additional AI contracts in the quarter without requiring additional capital of its own. With offtake-linked prepayment funding half the build and project debt funding much of the balance, payback on equity compresses towards ten months the same order as the large offtake deals above.
We are willing to state the conclusion bluntly: an asset that returns its gross cost in ten to twenty-six months, contracted take-or-pay, against a customer base bidding at premiums for four-year-old hardware, is not a capital misallocation. Where, specifically, is the bubble?
4. But are the offtakers making money?
The last line of the bear argument is that the counterparties signing these contracts are the ones absorbing the loss. The disclosed unit economics say otherwise.
The relevant metric is revenue per megawatt of compute, set against the cost of that megawatt. The base cost of running a megawatt of AI compute is around $10–15 million a year; frontier-model revenue on that megawatt has reached as high as $50 million in Anthropic's case. On our own assumptions, a gigawatt of frontier-class compute running a current-generation model generates $70–100 billion of annual revenue for the lab operating it, consistent with SemiAnalysis's framing that a gigawatt currently produces roughly $100 billion of revenue a year for about five years before replacement and with the $60 billion per gigawatt annualised monetisation rate used in recent a16z work.
Against $50 million per megawatt of contracted cost at the very top of the rental market, a lab monetising at $70–100 billion per gigawatt is comfortably profitable on the contract — and that is before considering that enterprise API revenue carries gross margins we assume close to 90%.
5. Is there any differentiation left between open and closed?
The final objection: if open-weight models converge on the frontier, price-performance collapses and the whole return stack goes with it. We disagree, on three grounds.
The frontier lead has been persistent, and open weights are partly derived from it
The gap has not closed. Epoch AI finds that since January 2026 the most capable open-weight models have lagged frontier closed models by an average of four months, or eight ECI points wider than the three-month average it measured between January 2023 and October 2025. On a China-specific cut, open-weight models have lagged US frontier models by an average of seven months since 2023. Call it roughly two quarters, and note that it has been widening rather than narrowing.
More importantly, a meaningful part of open-weight capability is derived from the frontier rather than developed independently. Anthropic disclosed industrial-scale distillation campaigns by three Chinese labs that generated more than 16 million exchanges through roughly 24,000 fraudulent accounts to extract reasoning, coding and agentic capability, and both Anthropic and OpenAI now treat this extraction as a national-security matter. A follower whose training signal is the leader's output cannot, by construction, overtake the leader.
Recursive self-improvement is now observable, not theoretical
We are firm believers in recursive self-improvement, and it has moved from thesis to disclosure. GPT-6 Astra is OpenAI's first model built at scale with substantial involvement from previous-generation models in the training process using AI to train AI, which OpenAI's research leadership credits with producing a deeper and more robust world model, at the cost of a more complex training pipeline. Industry chatter is consistent on a further point we cannot source publicly: both leading labs are understood to hold at least one more capable internal version beyond their latest public release. If that is right, the published frontier understates the real one, and the measured open-source lag is flattering to open source.
Both of the above are scaling laws continuing to work
The mechanism underneath is unchanged: bigger training clusters produce better models. GPT-6 Astra was trained on approximately 100,000-plus Nvidia Grace Blackwell NVLink72 systems, with Huang flagging 400,000 GPUs coming online next. Greg Brockman noting it was the first run trained on more than 100,000 GPUs, and the run itself executed at Oracle's Abilene site. The next frontier run is being sized at 300,000-plus GPUs on Vera Rubin silicon, a generational step up on a three-fold larger cluster.
What this means for pricing
Not all intelligence is the same, and the frontier will command a premium. A persistent capability lead, partially self-reinforcing through RSI, supports defensible margins at the frontier labs. The premium is for the marginal capability, and the marginal capability is where enterprise value gets created.
Open source will take the lion's share of total tokens generated. The blended price per million tokens will therefore keep falling. This is what the token expenditure price series actually shows — a mix shift, not frontier price compression and it is the most commonly misread chart in the sector. A falling blended average is fully consistent with a rising frontier price.
Every token costs the same to produce, whoever produces it. As Gavin Baker put it recently, a token is a token: it takes the same flops, memory and watts no matter which model produces it. Open-source share gains therefore redistribute margin into the compute layer rather than destroying demand — which is precisely why Nvidia supports open source, since more models mean more aggregate compute demand than one or two labs at 90% margins.
What would change our view
We would revisit this thesis on evidence, not narrative. Specifically: rental rates on prior-generation silicon rolling over on a sustained basis; renewal pricing turning to discounts rather than premiums; contract ACV per megawatt compressing back towards the $12 million 2026 base; prepayment coverage of capex falling materially; or algorithmic efficiency displacing brute-force scaling to the point that cluster size ceases to determine capability. None of these is present in current disclosure. The forward curve on the CME contract, once it lists, will be the cleanest single place to watch for the first of them.
Conclusion
The bubble case is an argument about the future presented as a description of the present. Every observable in the present: rental rates, renewal premiums on four-year-old chips, contract values per megawatt, disclosed payback periods, prepayment coverage, lab unit economics points the other way. Compute is scarce, it is clearing at record prices, capital committed to it is being returned inside two years, and the counterparties signing the offtake are now profitable on it.
Our positioning follows from the arithmetic rather than from enthusiasm. We remain constructive on the compute layer where the margin accrues regardless of whether the winning token is open or closed and on the frontier labs' ability to defend premium pricing for premium capability. We will change our minds when the data does.
Sources: Oracle Q1 FY27 earnings call (10 September 2026); CoreWeave Q2 2026 earnings call and subsequent disclosures; Nebius Q2 2026 results and 10-Q; SpaceX IPO prospectus and subsequent cloud service agreement filings; Silicon Data rental price indices (SDH100RT, SDH200RT, SDB200RT); Epoch AI Capabilities Index; Nvidia Q2 FY27 commentary; a16z and Invest Like the Best podcast interviews with Gavin Baker; SemiAnalysis commentary. Payback and per-megawatt figures are LC Equity Research derivations from disclosed contract values and capacity, not company-reported metrics.
.png)
.png)
.jpg)