You are negotiating the wrong term for your AI contract

Why this, now

This started with a client question, and it is the version I now get most often. They had a term sheet for AI compute in front of them, several hundred million dollars across four years, and they wanted to know whether the rate was competitive.

It was. However, that turned out not to be the useful thing to know.

The reason to write it up now is that the market has changed underneath the contracts being signed against it. For three years, AI compute procurement was a scarcity problem: you took what you could get, on the terms offered, and the only real question was whether you got an allocation at all. The supply position has since become far harder to read, and the negotiating posture most buyers learned during the shortage is now costing them money.

Duration compounds it. Agreements signed this year run three to five years, spanning roughly two and a half accelerator generations and a meaningful shift in how much of the world’s AI capacity is available on an open market at all — by which I mean a market where any buyer can purchase from several sellers at a price they can discover. Almost none of that risk is addressed by the rate.

The rate is the one term you cannot get right

Start with a fact that ought to be impossible in a functioning market. In September 2026, renting one H100 for one hour cost between $1.49 and $10.98 depending on where you bought it — Vast.ai’s peer-to-peer floor at the bottom, Google Cloud’s a3-highgpu-8g list price at the top, for the same silicon.¹ Cross-provider medians land in between and disagree with each other: AIMultiple tracked $3.38 in August 2026, the Mercatus index $3.84 across 27 providers in September.²

Now the direction. The expected pattern for a three-year-old chip is steady decline as newer silicon takes the premium workloads. That is not what happened. Bank of America put H100 spot at $2.77–$2.80 an hour in August 2026 — up 2% on the month and up 33% on the year. A100 pricing, on a chip two generations old, was up 17% year on year.²

Anyone who signed a multi-year fixed-rate commitment in 2024, when H100 rentals were above $7 an hour, is now paying well above the market median. Anyone who deliberately stayed on spot through 2026 expecting further declines paid a third more than they budgeted. Both bets were reasonable when made, and both are wrong in hindsight.

The reason is that the binding constraint moved, and it did not move to where anyone was looking. It is no longer GPU die output. Morgan Stanley described DRAM as having become the principal bottleneck in the AI buildout, with memory prices up roughly six-fold over the prior year as manufacturers redirected HBM and DRAM toward data-center parts.³ Nvidia has been unusually plain about the consequence: CFO Colette Kress told analysts in November 2025 that “the clouds are sold out,” and by August 2026 Jensen Huang was telling investors the company had “supply for 70% growth, but demand is much higher.” Kress added that supply would remain a bottleneck at least through the end of fiscal 2028.²

The same tightness shows up in captive silicon, and the buyers are saying so on the record. In April 2026 Anthropic committed more than $100bn to AWS over ten years and secured up to 5 GW of capacity spanning Trainium2 through Trainium4, with an option on later generations. Andy Jassy’s stated reason: the silicon is “in such hot demand.”⁴ Hyperscaler list prices, meanwhile, move slowest of all, because they are set through multi-year commercial agreements rather than marketplace clearing: AWS cut P5 pricing 44% in June 2025, and the resulting $6.88 an hour is still roughly 2.7 times a marketplace rate.⁵

So the honest summary is not “prices fell.” It is that the same chip traded across a seven-fold range in a single month, moved in opposite directions depending on which tier you bought, and did so because of a constraint — memory — that was not the one the market had been watching. You cannot win this negotiation on price. What you can do is decide in advance what happens when the price you agreed turns out to be wrong. That is a question about terms.

Nobody agrees what a unit of compute is

A surprising number of capacity agreements are denominated in something like “H100-equivalents,” and almost none say equivalent on what.

Training and inference are limited by different parts of the chip. A training run is compute-bound: it scales with low-precision operations per second. Inference decode is memory-bandwidth-bound: it scales with how fast weights can be pulled from high-bandwidth memory. So there are two defensible ways to normalize an accelerator against an H100, measuring different physical quantities:

Training basis — FP8 dense throughput in PFLOPS, against the H100 SXM5 reference of 1.979 PFLOPS.

Inference basis — HBM bandwidth in TB/s, against the H100 SXM5 reference of 3.35 TB/s.

Same chips, same reference, two answers. Figures below are from our capacity model.⁶

Take the H200, the cleanest case. It scores exactly 1.00 on the training basis — not approximately, exactly — because it carries the same GH100 compute die as the H100 and delivers identical FP8 dense throughput. Nvidia changed the memory: 141 GB of HBM3e at 4.8 TB/s against the H100’s 80 GB at 3.35.⁵ On a training-denominated contract the H200 is worth nothing more than an H100. On an inference-denominated one it is worth 43 percent more. Same part, same vendor, and the entire value of the upgrade lives or dies on a definition.

Now the four accelerators that land on exactly 2.53 on the training basis. GB300, Rubin NVL144, MI355X and Maia 200 are indistinguishable on that measure. On the inference basis they are 2.39, 3.88, 2.39 and 2.09 — an 86 percent spread between top and bottom, across assets a training-denominated contract would treat as interchangeable.

So an agreement for “the equivalent of 5,000 H100s” can be satisfied with materially different amounts of what you actually need, entirely legitimately, depending on an unstated convention.

The silicon designers have already picked a side. Setting out four MTIA generations in March 2026, Meta stated the premise plainly — HBM bandwidth is the most important factor affecting generative inference — and doubled bandwidth between MTIA 400 and 450 on that basis. Its stated strategy is “inference first,” on the explicit reasoning that mainstream GPUs are built for large-scale pre-training and then applied, often less cost-effectively, to inference.⁷ Meta sells no compute externally, so this is evidence about where the physics binds rather than about anyone’s contract. But in the term sheets, capacity is still counted on the training measure. That gap is an Aureak observation, not a published finding — and it is worth checking in your own agreements.

Captive silicon is becoming purchasable, but not tradable

Four internal accelerator programs are now at scale: Google’s TPU at v7; AWS Trainium in its third generation; Microsoft’s Maia, with the 200 announced in January 2026 and deployed in US Central⁸; and Meta’s MTIA, developed with Broadcom, with four further generations deploying across 2026 and 2027.⁷ The standard framing is Nvidia dependency. The contracting consequence gets less attention.

Until recently the position was simple: this capacity was captive. You rented it inside the owner’s cloud or not at all. That is changing, and it would be wrong to write as though it were not.

Google has crossed the line. On 18 May 2026 it formed a joint venture with Blackstone — an initial $5bn equity commitment, first 500 MW targeted for 2027 — to offer TPU capacity as a service alongside Google Cloud, with Google supplying hardware, software and services.⁹ On its July earnings call Google disclosed revenue from TPU systems delivered into customer data centers for the first time. For an asset strictly internal for its entire history, that is a genuine break.

Amazon is talking about it. Amazon confirmed in April 2026 that it is exploring direct sales of Trainium racks to third-party data centers. As of mid-2026 these were early discussions with no signed deal.¹⁰ Microsoft’s Maia and Meta’s MTIA remain fully captive, with no announced external path.

Purchasable is not tradable, and the distinction is what matters for a contract. By purchasable I mean capacity you can buy as hardware and operate in a facility the owner does not control — which TPUs now are, and Maia and MTIA are not. Tradable would mean something further: that having bought it, you could resell the allocation, price it against an independent published rate, or swap the commitment for another vendor’s silicon. None of those exist for any captive accelerator. The Blackstone vehicle is a controlled joint venture, not a spot market.

So the purchasable share of global AI capacity is rising while the tradable share falls. Your realistic set of alternatives at renewal is narrowing, fastest in large-scale inference where demand grows quickest. Substitution rights are becoming more valuable and harder to obtain at the same time.

The terms that decide the outcome

The unit of commitment. The highest-leverage term in the document, routinely treated as a drafting detail. A commitment in dollars is worth far more than one in GPU-hours, which beats one tied to a named configuration. Some providers offer portfolio draw-down — the right to spend across the vendor’s range rather than one SKU.¹¹ It is worth more than several points of discount, because it converts a bet on one generation into a bet on one vendor.

Technology refresh rights. Can you migrate to the next generation mid-term, on what notice, at what price? Without this, a multi-year commitment obliges you to keep running depreciating silicon while the market moves past it. Reserving a single generation for several years buys supply certainty and sells away suitability.¹²

The definition of the deliverable. Name the measure. If the contract says equivalents, it should say equivalent on FP8 dense throughput or on HBM bandwidth, with reference figures written in rather than left to a later argument.

Delivery date and the remedy for missing it. Energized capacity on a date is the scarce good. Yet the remedy for late delivery is usually service credits against the affected tranche — a rounding error next to the cost of a delayed launch. If delivery is what you are buying, the remedy should look like it.

Repricing and benchmarking. A periodic reprice to market, or a most-favored-nation clause against comparable customers, is rare and valuable. Against a chip that traded across a seven-fold range in one month, a five-year fixed rate with no reprice mechanism is a position, not a procurement outcome.

Shape rights. Can you burst above the reservation, and at what rate? Can you move the commitment between regions? Capacity you cannot shape is worth materially less, and the difference rarely appears in the rate comparison.

Change of control. Most agreements prohibit assignment without consent, and many allow pricing, credits, reserved allocation and committed-use discounts to be adjusted on a change of control.¹³ For a company that might be acquired, that converts a favorable contract into a negotiating liability at the worst moment.

The exit. Termination for convenience, wind-down rights, treatment of prepaid amounts. The standard structure is a minimum commitment, an overage rate, a capacity service level agreement and an exit clause¹¹ — and the exit is the least negotiated of the four.

What the other side is solving for

Two negotiations run in this market and they are easy to conflate. The first is between a chip vendor and an infrastructure operator: Nvidia or AMD sells accelerators to a neocloud or hyperscaler, which carries the asset on its own balance sheet. The second — the one this essay is about — is between that operator and you.

The operator typically finances the hardware it bought in the first negotiation against the contracts it signs in the second. Your contract is its collateral. What it can borrow depends on duration, fixedness of price, and absence of exits — precisely the three things you are trying to soften. CoreWeave is the clearest illustration: a $99.4bn revenue backlog at 31 March 2026, to be delivered from 49 data centers, substantially all of them leased rather than owned.¹⁴ That backlog is not a revenue forecast. It is the security package behind the leases.

The scale of that dependency is worth seeing. BloombergNEF tracked more than $100bn of hyperscaler leases for neocloud capacity in the six months to March 2026.¹⁵ Even the largest buyers in the world are contracting this way rather than building everything themselves, which is why the terms in these agreements have started to matter far beyond the parties signing them.

So every flexibility you win genuinely reduces the financeability of the contract. Flexibility is not free and you should not expect it to be given away. The error is not paying for it — the error is spending the whole negotiating budget on the rate, which the market will re-price for you anyway, then accepting standard terms on everything it will not.

Where this market goes

Standardization will arrive from the financing side, not the buying side. By standardization I mean convergence on common definitions — an agreed unit of capacity, a common service level, a standard delivery remedy — so one contract can be compared against another without reading both in full. Buyers have wanted that for three years and have not produced it, because every workload differs. Lenders have a stronger motive: you cannot securitize, syndicate or price a book of bespoke agreements. Watch the credit documentation, not the procurement templates.

And the market bifurcates. Merchant silicon moves toward a commodity market — interchangeable units, a visible reference price, short tenors, low switching costs. Captive accelerators go the other way: purchasable, but illiquid, unbenchmarked and bundled into a platform relationship. Buyers will increasingly choose not between vendors but between market structures. Both are defensible. They are different risks, and they should not be evaluated with the same spreadsheet.

What this means for your decision

If you are buying compute, seven questions, in this order, before discussing rate at all. What exactly am I committing to — dollars, hours, or a named configuration? On what measure is capacity counted, and is that measure written into the agreement? What happens in month eighteen when a better accelerator ships? What is the remedy if capacity is late, and does it bear any relation to what late capacity costs me? Can I reprice to market during the term, or benchmark against comparable customers? Can I burst above the reservation and move it between regions? And what happens to this contract if we are acquired? Only then the rate, with the explicit intention of trading some of it for the other seven answers.

If you are underwriting a provider, read the book of contracts as a credit instrument rather than a revenue forecast. Weighted average duration, the proportion of revenue denominated in dollars versus named SKUs, and the number of counterparties holding refresh or substitution rights all tell you how much of that backlog is genuinely contracted. A book heavy in named-SKU commitments looks strong until a generation turns.

The buyers who look smart in 2028 will not be the ones who got the best price in 2026. Prices move in both directions and nobody has forecast them accurately yet. Buyers who look smart will be the ones who kept the ability to change their mind, in a market steadily withdrawing that option from everyone who did not negotiate for it.

References

1. H100 on-demand rates spanning $1.49/hr (Vast.ai, via CloudZero) to $10.98/hr (Google Cloud a3-highgpu-8g, via Thunder Compute) per GPU, September 2026. Hyperscaler band reported as $6.88–$10.98/hr. https://shattered.io/h100-h200-b200-cloud-gpu-pricing-2026/ ↩
2. AIMultiple Cloud GPU Rental Price Index, 25 August 2026: median H100 $3.38/hr, down from $7+/hr in early 2024. Mercatus GPU Index, 13 September 2026: $3.84/hr across 27 providers. Bank of America (Vivek Arya), August 2026: H100 spot $2.77–$2.80/hr, up 2% month on month and 33% year on year; A100 $1.65/hr, up 17% year on year. Nvidia FQ3 2026 call, 19 November 2025 (Kress) and FQ2 2027 call, 26 August 2026 (Huang, Kress). All as reported, September 2026. https://shattered.io/h100-h200-b200-cloud-gpu-pricing-2026/ ↩
3. Morgan Stanley, as reported by Reuters, 3 June 2026: memory chip prices up roughly six-fold over the prior year; DRAM described as the principal bottleneck in the AI infrastructure buildout. https://shattered.io/h100-h200-b200-cloud-gpu-pricing-2026/ ↩
4. Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute”, 20 April 2026. Commitment of more than $100bn to AWS over ten years, securing up to 5 GW of capacity spanning Graviton and Trainium2 through Trainium4, with an option to purchase future generations; nearly 1 GW of Trainium2 and Trainium3 capacity online by end-2026. Andy Jassy quoted on custom silicon being “in such hot demand.” https://www.anthropic.com/news/anthropic-amazon-compute ↩
5. Spheron H100 price tracker, rates as at 14 June 2026, page updated 8 September 2026. AWS cut P5 H100 list pricing approximately 44% in June 2025; the resulting $6.88/GPU-hr remains roughly 2.7x Spheron’s SXM5 on-demand rate. Also states that the H200 uses the same GH100 compute die as the H100, upgrading memory from 80 GB HBM3 to 141 GB HBM3e at 4.8 TB/s, delivering 37–90% faster inference on 70B+ models. Spheron is a GPU marketplace; its own rates are vendor-published. https://www.spheron.network/blog/nvidia-h100-news-2026/ ↩
6. Aureak model estimate. H100-equivalents from the Aureak capacity model, v26, September 2026. Training basis normalizes FP8 dense throughput against the H100 SXM5 reference of 1.979 PFLOPS; inference basis normalizes HBM bandwidth against 3.35 TB/s. Accelerator specifications compiled from vendor datasheets and conference disclosures; figures for unreleased parts are Aureak estimates, not observed data. ↩
7. Meta AI, “Four MTIA Chips in Two Years: Scaling AI Experiences for Billions”, 11 March 2026. MTIA 300, 400, 450 and 500, developed with Broadcom, deploying across 2026 and 2027. States that HBM bandwidth is the most important factor affecting GenAI inference performance, and that bandwidth was doubled from MTIA 400 to 450 on that basis. Sets out an explicit “inference first” strategy. https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions/ ↩
8. Microsoft, “Maia 200: the AI accelerator built for inference”, 26 January 2026. TSMC 3nm, 140bn transistors, 750 W TDP, 216 GB HBM3E at 7 TB/s, 272 MB on-chip SRAM, approximately 10 PFLOPS FP4 and 5 PFLOPS FP8; deployed in the US Central region. https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/ ↩
9. Blackstone, “Blackstone Announces Joint Venture with Google to Create New TPU Cloud”, 18 May 2026. Initial $5bn Blackstone equity commitment; first 500 MW targeted for 2027; Google to supply hardware including TPUs, plus software and services. https://www.blackstone.com/news/press/blackstone-announces-joint-venture-with-google-to-create-new-tpu-cloud/ ↩
10. Amazon confirmed in April 2026 that it is exploring direct sales of Trainium racks to third-party data centers; reported as early discussions with no signed agreement as at June 2026. https://qz.com/amazon-trainium-ai-chips-external-data-centers-061826 ↩
11. Typical GPU capacity reservation structures comprise a minimum commitment, an overage rate, a capacity guarantee service level agreement and an exit clause; some providers offer draw-down of the commitment across the vendor portfolio rather than a single configuration. As at 2026. https://www.spheron.network/blog/gpu-cluster-reservation-contract-negotiation-2026/ ↩
12. A customer reserving a single hardware generation for a multi-year term gains supply certainty while accepting the risk that capacity becomes unsuitable, uncompetitive or unnecessary as models and demand change. As at 2026. https://aitechmodel.com/compute-offtake-agreements-for-ai-startups-what-changed-and-why-it-matters-now/ ↩
13. Mayer Brown, “Portability of AI Compute Infrastructure in AI Acquisitions”, June 2026. AI compute agreements commonly prohibit assignment without the provider’s prior written consent and may permit adjustment of pricing, credits, reserved capacity and committed-use discounts upon a change of control. https://www.mayerbrown.com/en/insights/publications/2026/06/portability-of-ai-compute-infrastructure-in-ai-acquisitions ↩
14. CoreWeave, Inc., Q1 2026 earnings press release (Form 8-K). Revenue backlog of $99.4bn as at 31 March 2026, to be delivered from 49 data centers, substantially all leased from third parties. https://www.sec.gov/Archives/edgar/data/0001769628/000176962826000220/coreweave1q26earningspress.htm ↩
15. BloombergNEF, “AI Data Center Build Advances at Full Speed: Five Things to Know”. Hyperscaler leases for neocloud-provided compute capacity in excess of $100bn over the six months to March 2026. https://about.bnef.com/insights/data-centers/ai-data-center-build-advances-at-full-speed-five-things-to-know/ ↩