Executors reporting AI spend above budget
59%
Survey of 500 large-company CXOs, Sept. 2026 fieldwork, published Nov. 12, 2026
Average overshoot where AI budgets broke
24%
21% of organizations overshot their AI budget by more than 30%
Anthropic IPO value at the top of range
$1.41T
Priced Nov. 10, 2026 at $70 to $78 a share, fully diluted
Anthropic contracted compute obligations
$417B
Plus $125B to $150B of ordered compute for 2028 to 2031, per its S-1
The Demand Signal
The Cost Shock Is Landing on the Buyer's Side of the Ledger
Capgemini Research Institute's November 2026 report, based on a Sept. 2026 survey of 500 CXO-level executives at organizations with more than $1 billion of revenue across 10 countries and 13 industries, found that 59% ran AI spending materially above plan over the prior 12 months, with the average overrun at 24% and 21% overshooting by more than 30%. Two-thirds of the organizations that broke budget reported negative financial consequences, including lower free cash flow or reduced discretionary spending. The shock is a demand-side repricing of AI, not a preference for open source, and it landed two days after Anthropic set its IPO price range.
- Nearly three in five organizations plan to expand their use of open-weight models over the next 12 months, and more than three in five expect to adopt a multi-model strategy.
- 53% tightened expenditure controls after unexpected AI spend, and only 56% are fully confident in forecasting how AI costs evolve as adoption scales.
- Six in ten lack a complete view of AI costs, and one in two approve AI spending without adequate cost visibility.
- One company ran a $500 million AI bill in a single month after failing to set usage caps; among firms that track spend systematically, 20% of AI expenditure does not ultimately create business value.
- AT&T is cited as moving from roughly 40% to 70% of its AI workloads onto open models within a year.
The Price of a Token
Open Weights Deliver Near-Frontier Capability at a Fraction of the Price
| Route | Price per million tokens | Where it wins |
|---|---|---|
| Self-hosted open weights on owned GPUs | About $1 | Break-even near 5 to 10 million tokens a day, provided utilization is high |
| Open-weight models via API routers | $0.13 average open-source price | Already 19% of enterprise AI tokens; 72% of tokens routed on OpenRouter |
| Closed frontier, blended average | $0.70 | Still 63% of enterprise AI spend, so capability and tooling still command the money |
| Closed frontier, premium tier | About $50 | The price point enterprises are explicitly trying to avoid |
The substitution is not a capability bet. The gap between the best closed and the best open-weight model on Arena narrowed from 15% to about 3%, while open-weight tokens cost 87% less on average and deliver roughly 90% of capability at about 16% of the price. Routing, not switching, is the real mechanism: Capgemini finds that model routing cuts AI costs by 27% to 85% at near-frontier quality, which is why buyers are adding open-weight capacity to a portfolio rather than cancelling their frontier contracts.
- Open-weight tokens cost 87% less on average on OpenRouter, and blended closed-model pricing has fallen 24% in a year, so both sides of the market are defating.
- Blended closed models still average $1.86 per million tokens against $0.23 for open-weight models in the Capgemini sample, a gap wide enough to fund routing decisions.
- Nvidia publishes and optimizes its own open-weight model family for Blackwell-class inference, so the hardware vendor is also a model vendor in the substitution trade.
The Hidden Divergence
Tokens Are Migrating to Open Weights, and Dollars Are Not
Open-weight share of enterprise AI: tokens up, spend down
Menlo Ventures, 2026 State of Enterprise AI, Nov. 3, 2026. Open-weight token share rose 73% year over year while open-weight spend share fell.
Unit: % of enterprise AI
Open-weight tokens, Sep. 2025
11%
Open-weight tokens, Sep. 2026
19%
Open-weight spend, Sep. 2025
6%
Open-weight spend, Sep. 2026
4%
Enterprise AI deployment spending reached $176.8 billion by Sept. 2026, up from $33 billion a year earlier, and frontier labs now take 69% of it, or $122.7 billion, against 50% last year. Inside that total, Anthropic's share of enterprise LLM API spending rose from 40% to 55% and OpenAI's from 27% to 31%, leaving the two of them at 86% of API spending, while Alphabet's fell from 21% to 9%. The cost shock has not yet cost frontier labs revenue, because falling token prices are being more than offset by volume growth, and it is precisely that revenue, 80% consumption-based, that buyers are now underwriting at 21 times sales.
The First Price Discovery
Anthropic's Listing Is Where the Cost Shock Meets a Public Market
| Anthropic | OpenAI | |
|---|---|---|
| Valuation mark | Up to $1.41T at $70 to $78 a share, Nov. 10, 2026 | $1.25T private round, Nov. 6, 2026 |
| Latest annualized revenue | About $67B, from roughly $16.9B in the third quarter | About $50B in Sept. 2026 |
| Multiple on the current run rate | 21x | 25x |
| Company-indicated year-end run rate | $100B | $70B |
| Multiple on that target | 14x | 18x |
The revolt was visible before the range. Reuters reported on Nov. 5 that prospective investors viewed $1.5 trillion as the high end of expectations, were pushing for a more attractive valuation, and said Anthropic may need to offer a bigger discount to secure commitments; The Wall Street Journal reported the same skepticism after a sharp technology selloff, with some investors questioning whether the roadshow could finish in time for a November listing. Anthropic had also left October revenue blank in the updated filing, the one number buyers were waiting for. What the range settled at says buyers did not reject the business, they repriced the risk, and the repricing came from $1.5 trillion to $2 trillion down to $1.26 trillion to $1.41 trillion.
It's not like these companies are not going to be a part of the future. They're going to become public. Valuations have risen for a long time in the private market, and we'll probably see some pullback.
- $417B of contracted compute obligations, plus $125B to $150B of ordered capacity for 2028 to 2031, stand against $33.2B of nine-month revenue.
- The $104.0B nine-month net loss includes $94.3B of fair-value remeasurement on convertible notes, derivatives and SAFEs, so excluding those marks the loss is about $9.7B, and the optics were what buyers fought over.
- 80% of first-half revenue was consumption-based, $13.1B of $16.29B, so a 3.3 times annual fall in token prices lands directly on the revenue line.
- Growth is slowing even as it stays huge: year-over-year revenue growth of 1,304% in the first half eased to 1,206% for the nine months.
- 48% of first-half revenue ran through Amazon, Google Cloud and Microsoft Azure, and cloud partners take about 16 cents of every dollar Anthropic books through them, so the channel is also the cost line.
Anthropic's filing, in numbers
Revenue, nine months to Sept. 30, 2026
$33.2B
Up 1,206% year over year
Net loss, same period
$104.0B
Includes $94.3B of non-cash fair-value marks
Compute obligations
$417B
Plus $125B to $150B ordered for 2028 to 2031
Customers above $100,000 trailing revenue
11,900+
More than 1,900 above $1M and about 250 above $10M
Consumption-based revenue, first half 2026
80%
$13.1B of $16.29B; subscription revenue $3.2B
Supply Chain
The Money Moves Down the Stack, Not Out of It
Frontier-lab pricing power is the layer under pressure; the layers beneath it are not. NVIDIA reported $96.2 billion of revenue in its fiscal second quarter of 2027 with data-center revenue of $89.0 billion, up 117%, and a 75% gross margin. Broadcom posted $29.6 billion of revenue at a 69.1% gross margin in its fiscal third quarter of 2026, and Micron turned $41.5 billion of revenue at an 84.6% gross margin in its fiscal third quarter ended May 28, 2026, which shows memory pricing is not deflating alongside model pricing. Open-weight substitution relocates enterprise spending from model licenses to rented silicon, and that is why buyers chasing cheaper tokens are not cutting infrastructure orders.
| Layer | Who | Mechanism | Hard number |
|---|---|---|---|
| Accelerators | NVIDIA | Substituted tokens still consume GPU hours | Data-center revenue $89.0B, up 117%, Q2 fiscal 2027 |
| Custom silicon | Broadcom | Open-weight serving rewards tunable self-hosted hardware | $29.6B revenue at 69.1% gross margin, Q3 fiscal 2026 |
| Memory | Micron | Token volume, not token price, drives content per cluster | $41.5B revenue at 84.6% gross margin, FQ3 2026 |
| Cloud and marketplace tolls | Amazon, Alphabet, Microsoft | Anthropic, open weights and everyone else route through them | $110B Amazon commitment; $114.5B Google TPU commitment |
| Neocloud hosting | CoreWeave | Hosts sell the GPU hours that open-weight inference requires | $132.5B revenue backlog at Sept. 30, 2026 |
| Open-weight model supply | Alibaba | Qwen is sold with a revenue-share wrapper on top of open weights | Cloud revenue up 45%; AI product revenue triple-digit for a 12th straight quarter |
| Frontier labs | Anthropic and OpenAI, both private | Price per token deflates while committed compute does not | $417B of obligations against $33.2B of nine-month revenue |
- Seven of Anthropic's own suppliers agreed to buy about $26 billion of Anthropic stock at the IPO price: Alphabet, NVIDIA, Advanced Micro Devices, Broadcom, Micron, Microsoft and TSMC.
- Amazon holds $6.2 billion of convertible notes that could convert into up to 1.84 billion shares, with its stake capped at 12%, so the largest customer is now a capped equity holder in its own supplier's cost base.
- Amazon spent $77.3 billion on property and equipment in the September quarter and $215.2 billion over the trailing year, while Microsoft spent $39.4 billion in its September quarter, so the hosting build is unchanged by which model wins.
Horizons
What Moves First, and What Breaks Last
Over the next four weeks, three dated prints settle the argument with hard numbers rather than surveys.
- Nov. 17, 2026: NVIDIA reports fiscal third-quarter 2027 results, the first clean read on whether open-weight inference expands the compute bill or merely reshapes it.
- Nov. 19 to 20, 2026: Anthropic's Nasdaq debut, priced at $70 to $78 after valuation talks slid from $1.5 trillion to $2 trillion down to $1.26 trillion to $1.41 trillion.
- November: Alibaba's September-quarter result, where 45% cloud growth and a revenue-share wrapper on open weights must outrun a 75% capex increase that cut operating income 57% to RMB15.2 billion.
- The first post-listing quarter, when consumption revenue has to grow into a 21 times sales multiple after open-source token prices fell 3.3 times in twelve months.
Over one to three years, the structural question is where the deflation lands. If open-weight prices keep falling at multiples while capability stays within a few points of the frontier, the marginal AI workload becomes a commodity served by rented silicon: model vendors keep the capability premium on the hardest tasks and everyone else buys tokens at the open-weight clearing price. The $417 billion of contracted compute is what makes this a race rather than a glide path, because commitments do not reprice when token prices do. The likely end state is more AI spending at lower unit prices, which is excellent for the infrastructure layer and punishing for labs whose fixed commitments are sized to today's prices.
Where the substitution trade actually lands
- Data-center revenue of $89.0B, up 117% in Q2 fiscal 2027, shows open-weight substitution still clears through Nvidia silicon.
- Substitution moves enterprise spend from model licenses to rented GPU hours, supporting accelerator demand even as token prices fall 3.3 times a year.
- Near term: fiscal Q3 2027 results on Nov. 17, 2026 test whether token volume growth outruns inference price deflation.
- Q3 fiscal 2026 revenue of $29.6B at a 69.1% gross margin, plus a seat on Anthropic's $26B supplier placement, ties custom silicon to both frontier labs and open-weight serving.
- Open-weight workloads favor tunable self-hosted accelerators over pay-per-token APIs, which is exactly where custom silicon wins.
- Long term: the more inference moves in-house, the more of the $89.0B data-center bill shifts toward customized silicon.
- Fiscal Q3 2026 revenue of $41.5B at an 84.6% gross margin shows memory pricing is not deflating with model pricing.
- Token volume rather than token price drives memory content per cluster, and the 19% open-weight token share still needs the same hardware.
- Near term: memory pricing is the cleanest read on whether substitution is expanding volume, with a 3.3 times annual token price decline as the offset.
- A $110B compute commitment through April 2036 and $6.2B of convertible notes make Amazon both Anthropic's landlord and a capped 12% shareholder.
- Partners take about 16 cents of every dollar Anthropic books through them, so deflation in frontier token pricing compresses Amazon's toll as well.
- Near term: Q3 2026 capex of $77.3B and trailing capex of $215.2B show the hosting build is unchanged by which model layer wins.
- A $114.5B TPU commitment through 2033 and a slot in Anthropic's $26B placement lock in demand.
- Alphabet is the clearest casualty of budget-driven substitution: its share of enterprise LLM API spend fell from 21% to 9% in a year while Anthropic and OpenAI took 86%.
- Long term: a broad open-weight catalog can recapture hosting and platform revenue, but at lower margin per token.
- Roughly 48% of Anthropic's first-half revenue ran through the three large clouds, and Microsoft also operates a Foundry catalog of more than 11,000 models.
- Microsoft monetizes the routing decision itself, whether the winning model is a closed frontier API, Qwen or a self-hosted checkpoint.
- Near term: Q1 fiscal 2027 capex of $39.4B and trailing operating cash flow of $195.9B fund hosting capacity regardless of the model winner.
- Revenue backlog of $132.5B at Sept. 30, 2026 against $3.66B of quarterly revenue means the hosted-GPU queue is long regardless of which model vendors win.
- Its own analysis puts self-hosted open weights near $1 per million tokens versus about $50 for premium closed models, making substitution a volume thesis for GPU hosting.
- Risk: a $709M quarterly net loss and $41.6B of recourse debt leave little room for utilization shortfalls if token prices fall faster than volume grows.
- Cloud revenue of RMB48.4B, up 45%, with AI product revenue growing triple-digit for a 12th straight quarter, puts open-weight supply on a monetization path frontier labs lack.
- The pivot to charging large users of Qwen for a share of the revenue their models generate converts free weights into an operating business.
- Risk: capex of RMB67.7B, up 75%, drove operating income down 57% to RMB15.2B, so monetization must land before depreciation does.
