I Stopped Paying $20 a Month. Can CoreWeave Survive People Like Me?
A couple of years ago I paid OpenAI $20 a month for ChatGPT. Today I pay nothing. Free tiers of several frontier models, plus open-source models like DeepSeek, get my work done.
Last week, CoreWeave, which rents out the chips that run many of these models, had its worst week in about eleven months. Its stock fell about 25%, and one of its bonds briefly yielded over 16%. Wall Street is arguing about whether CoreWeave can pay its lenders. Almost no one is asking whether people like me will pay enough for its customers to renew their contracts with the likes of CoreWeave.
A larger question to ask is that for Founders, how the CoreWeave situation may affect capital raising for startups.
My switch is that debate (Bear and Bull camp) in miniature. It raises four questions:
Will most end users pay at all? Or was my $20 a month the exception?
Are free tiers a promotion or the new normal? If subsidized tokens become the norm, someone else pays for the chips.
Can paying consumers and enterprises cover the hyperscalers' bills? Microsoft and others sign five-year, fixed-price, must-pay contracts with data center operators like CoreWeave.
Can the data center cover its own lease? CoreWeave's average lease runs about 12 years. Its customer contracts run about 5.
Underneath all four is one question: are we funding long-term commitments with short-term revenue, while the price of a token falls and open-source models take a growing share of the work?
My short answer: so far, soaring usage has outrun falling prices, but the paying base is narrow and cheaper open models are taking a growing share of the work. If premium models keep a clear lead, the system holds. If they lose it, the damage shows up when the five-year contracts come up for renewal, around 2028 to 2031. This is a slow fuse, not a fire today.
CoreWeave’s bad week, in numbers
Stock. Down 15.5% on October 14 and 11.9% on October 16, closing at $61.39. About 25% down for the week. A 52-week low near $58 to $59 on October 20.
Bonds. The October 2031 bond briefly yielded over 16%. Forbes reported about $4 billion of CoreWeave debt below 80 cents on the dollar. For comparison, Fortune cited Oracle's 10 to 12 year debt at roughly 8% to 8.5%.
Balance sheet. Debt rose from $14.6 billion to $51.6 billion in a year. Second-quarter interest was $640 million, about a quarter of revenue. Planned 2026 spending of $35 billion to $39 billion is about three times expected revenue.
Concentration. Three customers were about 72% of second-quarter revenue. Microsoft alone was about 67% of 2025 revenue.
Two camps
The two camps split cleanly. Bears (XTB, HSBC with a $39 target, Gartner on neoclouds) say CoreWeave is the canary. Bulls (Goldman Sachs at $106, Cantor at $176, JPMorgan, and CEO Michael Intrator) point to a $104 billion backlog and chips in short supply.
Both camps are arguing about whether CoreWeave can pay its lenders. Neither is asking whether its customers' customers, people like me, will pay enough for the contracts to be renewed at the prices they currently command.
Long commitments, short revenue
Every link in the chain is committed for longer than the link that pays it. My cancellation sits at the very bottom, where the commitment is shortest.
Link in the chain | Who commits | How long |
Consumer subscription (my old $20 plan) | End user | 1 month, cancel anytime |
Enterprise seats and API spend | Businesses | Often annual or pay-as-you-go |
GPU capacity contract, must-pay | Microsoft, Meta and other hyperscalers | About 5 years on average |
GPUs, as depreciated on the books | CoreWeave | 6 years |
Data center lease, average time left | CoreWeave | About 12 years |
So CoreWeave signs 12-year leases to hold 6-year chips that earn money from 5-year contracts. Those contracts are paid by hyperscalers whose own customers can walk away in a month. CoreWeave's investor deck says cash left after one 5-year contract could provide additional cushion for about five more years of leases. That only works if the contracts renew at good prices.
The chip side of the math has not changed. At October 2026 rental rates, an Nvidia H100 at $2.82 an hour and a B200 at $5.93 an hour both cost roughly $0.26 to $0.46 per million tokens for a mid-size model, depending on how busy the chip is. The newer chip does twice the work at twice the rent, so cost per token barely moves.
The question is what happens on the other side: the price of a token, the number sold, and who pays for them. The sections below work through each.
The price of a frontier token, 2023 to 2026
The big price cut happened once, between GPT-4 in 2023 and GPT-4o in 2024. Since mid-2024, mainstream flagship prices have fallen only about 36%, to about $4 per million tokens, and the top tier has stayed between $20 and $30.
One thing this chart does not show: the price of a fixed level of capability falls far faster; Epoch AI estimates about 13 times a year since 2023.
At the list-price level, though, the decline for mainstream flagships has mostly stopped. The pressure now comes from customers moving work to cheaper models, shown in the open-source section below.
How many tokens: usage volume, 2024 to 2026
Volume has grown far faster than price has fallen. Google reported 9.7 trillion tokens a month in May 2024 and 3.2 quadrillion in May 2026, about 330 times more.
Other platforms point the same way: OpenRouter, a developer marketplace, grew from roughly 10 trillion tokens a year to over 100 trillion by mid-2025, per its study with a16z.
The second thing the chart shows is deceleration. Year-on-year growth fell from about 49 times to about 7 times. Volume still beats a 36% price drop easily, but usage growth tends to slow as a market matures, and the contracts run five years.
Price times volume: the implied revenue
Multiplying the two series gives an implied token value that grew about 211 times in two years, from roughly $0.7 billion to $154 billion a year. Volume rose about 330 times; price fell about 36%.
Implied annual value = tokens per month × 12 × (price per 1M tokens ÷ 1,000,000)
Year | Google tokens per month | Tokens per year | Workhorse frontier price, per 1M | Implied annual value |
2024 | 9.7 trillion | 116.4 trillion | $6.25 | about $0.73 billion |
2025 | 480 trillion | 5,760 trillion | $4.31 | about $24.8 billion |
2026 | 3,200 trillion | 38,400 trillion | $4.00 | about $153.6 billion |
Read this as a ceiling, not as revenue. Three caveats matter:
It is one company's volume. Google is the only lab that has published a comparable yearly series. Its count includes input and output tokens across all its products, including free ones. There is no verified industry-wide figure for frontier-model tokens.
Most of those tokens are not sold. They are served free inside Search, Gmail and the Gemini app, and are likely paid for by ads. The $154 billion is what the tokens would be worth at mainstream list price. That gap is the subsidy question in one number.
The mix shift cuts it further. On Vercel's gateway in September, open-weight models carried 71.5% of tokens but only 23% of spend. That implies an open-weight token earns about one-eighth of a closed one. At that mix, average revenue per token is about 37% of an all-frontier mix, which would cut $153.6 billion to roughly $57 billion.
Volume has beaten price so far by a wide margin. But the revenue that must cover five-year chip contracts depends on how many of those tokens anyone pays for, and at what model's price.
Open source's share of usage
On developer gateways, open-weight models went from a sliver of tokens to the clear majority in under a year, while still earning a minority of the money.
Platform | Earlier reading | Latest reading |
Chinese models under 2% of tokens about a year earlier; open models roughly a third of tokens by late 2025 (secondary reports of the a16z study) | Open-weight 69.1% of named token volume, June 2026 (Tom Tunguz) | |
7% of tokens, December 2025 | 71.5% of tokens and 23% of spend, September 2026 |
The gap between the two lines is the revenue problem. Every open-weight token still runs on a GPU somewhere, so chip demand holds. But at September's mix, an open-weight token earned roughly one-eighth of a closed one. Spend share is now rising too, from 5% to 23% in three months, led by pricier open models like Kimi K3. That money increasingly goes to whoever hosts the model, not to the labs whose premium tokens justify five-year chip contracts.
Who actually pays: free tiers, subscribers and enterprises
Most people who use AI do not pay for it, and the ones who do are increasingly businesses spreading their money across cheaper models.
Consumers: a small paying minority. OpenAI reports more than 900 million weekly ChatGPT users and more than 50 million consumer subscribers. That is roughly 1 paying user in 18, about 5.6%. These are company-reported figures, not audited. At $20 a month, a subscriber like my former self is worth $240 a year. Even if all 50 million paid exactly $20, that is about $12 billion a year, before Pro plans at higher prices.
Free tiers look like the norm, not a promotion. Google's 3.2 quadrillion monthly tokens are mostly served inside free products. AI Overviews alone reaches more than 2.5 billion monthly users. Those tokens are likely paid for by advertising, not by anyone buying tokens. The price-times-volume section above shows how large the gap is between what those tokens would cost at list price and what anyone pays for them.
Enterprises: paying more, but less loyal. On Vercel's AI gateway, September spend was about 13 times December's while the average price per token fell 47%. Spend on frontier models rose about tenfold; spend on open-weight models rose more than seventyfold. Almost one in three teams changed their most-used lab in a single month.
The bull case is real. Lab revenues are growing very fast. OpenAI's annualized revenue was reported above $25 billion early in 2026 and is now at an annual run-rate of $40 billion, short of its own projected $70 billion.
So, can paid tiers cover the five-year commitments? For now, revenue growth says yes. But the paying base is narrow: a few percent of consumers and a set of enterprises that switch models monthly. Those customers are on 1-month to 1-year terms. The data centers they ultimately fund are on 5-year terms with their customers but 12-year lease terms with their own landlords. The mismatch is real.
What this could mean for startup financing
The opening of this piece asked how the CoreWeave situation may affect capital raising for startups. It is too early to measure, because the evidence so far is one bad week in public markets. What follows separates what has been reported from what is my inference.
Where the money is today
Venture funding is still strong and heavily concentrated in AI. Crunchbase News's Q3 2026 report, as summarized by several secondary outlets (I have not read the original report), puts global venture funding at about $159 billion, of which AI-driven startups took about $102 billion, or 64%. Total funding was down about 25% from the second quarter but up about 53% from a year earlier. A record 27 companies raised rounds of $1 billion or more, together close to a third of the quarter's total, and the largest rounds included infrastructure companies such as Crusoe and Nscale.
This matters because the investors funding AI infrastructure are, to a large degree, the same investors funding everyone else. A fixed-income manager quoted in a Reuters analysis of the selloff made a similar point about public markets: the exposure is held in the same portfolios by the same kinds of investors. If those investors reprice risk in one corner, it can spread.
Five ways the pressure could reach founders (analysis, not forecast)
Infrastructure and neocloud startups are exposed first. Reuters reported that Gartner expects 80% of neocloud providers to stop operating independently by 2030. Bloomberg, as relayed by secondary outlets, reported investor pushback on Nscale's proposed $30 billion IPO valuation, with some investors suggesting $16 billion to $19 billion. Founders building GPU capacity or data centers should expect lower valuations and harder questions.
Debt may get more expensive. Infrastructure startups often borrow against chips and customer contracts. When a bellwether's bonds yield over 16%, lenders tend to reprice similar risk.
Diligence moves from growth to durability. The questions in this article become investor questions: how many users pay, for how long, how concentrated the customers are, and how much revenue leaks to cheaper open-weight models. CoreWeave's three customers at about 72% of second-quarter revenue is the kind of concentration markets are now punishing.
Capital may stay concentrated. If investors turn selective on AI infrastructure, the money may go to fewer, larger names rather than flow back to the broad market. The median startup would then compete for a thinner share of a pool that looks large in aggregate. The opposite is also possible: a pullback from infrastructure could free capital for application-layer and non-AI companies.
Founders who rent compute may come out ahead, with a catch. If financing stress pushes neoclouds to discount capacity to fill it, startups that buy compute could see lower prices. That is speculation. The catch is counterparty risk: a startup with reserved capacity at a stressed provider could face renegotiation or disruption.
What would change this view
The bull case has real support. Overall venture funding is well above a year ago. CoreWeave's backlog was reported at $104 billion in June, and Nvidia has reportedly agreed to buy CoreWeave's unsold capacity through April 2032. Credit analysts have also disagreed about how serious the funding problem is. Signals worth watching are CoreWeave's bond yields, the outcome of the Nscale IPO, the terms on new infrastructure debt rounds, and whether Crunchbase's Q4 2026 figures show AI's share of venture dollars holding or falling.
Practical takeaways for founders
Raise before you need to, and avoid tying your runway to a single financing source.
Match commitments to revenue. Avoid long compute commitments against month-to-month revenue. It is the same mismatch this article describes in the chain above.
Know your numbers by model. Be ready to show gross margin, paying versus free users, and how easily customers could switch to a cheaper model.
Check your compute vendor. Look at its financial health and at your contract's termination and transition terms.
My judgment, not a forecast: the likelier near-term effect is tougher terms and sharper questions for AI infrastructure founders, not a freeze across venture. A wider chill would require the credit stress to spread beyond neoclouds.
Conclusion: two risks at two speeds
Yes, we are funding long-term commitments with short-term revenue, and it works only as long as token volume keeps outrunning the fall in revenue per token.
The fast risk is financing. With bond yields near 15% and interest about a quarter of revenue, new debt gets expensive now. The bond market is pricing this today.
The slow risk is renewal. If revenue per chip-hour falls short, Microsoft, Meta and others will ask for shorter or cheaper terms when contracts expire, around 2028 to 2031. CoreWeave's 12-year leases do not shorten.
My own switch from $20 a month to free is the bears' best argument in one line: the chips are rented on five-year terms, but the people at the end of the chain can stop paying in a month, and many never started. The bulls are probably right about the next two years. The renewal years will show who is right after that.
More Insights
Tony Zhang, Author, is Founder, Sbur, a 900+ Founders community. The community members get together often in-person or online to share critical insights, brainstorm solutions, make introductions, and help one another make things happen.
If you’re a Founder/Builder, join us: https://app.sbur.com/invite?group=36
Comments