The technology works, and adoption is real. Low-cost, open-weight (open-source AI) models get better every four months. They put frontier-level tools in every business's hands at a fraction of the price. Continue reading
At long last, we have Anthropic’s S-1. A first look into one of the main drivers of the AI boom or AI crash. So far, OpenAI has not filed audited results for the 2nd or 3rd quarter, so comparing audited results against rumors is specious and inaccurate. But for today, it is the best we have. Let’s play with what we got.
Google’s AI Overview says the coming AI crash is industry-wide. The Anthropic S-1 tells a dramatically different story. Also in the story are Meta, SpaceX, Microsoft, Google and Amazon, which have all made large AI investments but have other income sources to cover the debt.
Compare the same quarter. In Q2 2026, Anthropic booked about $11.6 billion in revenue, while OpenAI booked $6.7 billion and lost $12.3 billion, according to the Wall Street Journal. This is not an exact comparison. Both companies count cloud-partner sales. We also shouldn’t discount US government sales, or the administration tilting the scales with government contracts.
| Anthropic (S-1, audited) | OpenAI (press reports) | |
| Source | S-1, filed 10/26/2026 | WSJ (8/18), investor deck (Oct) |
| Q2 2026 revenue | ~$11.6B | $6.7B |
| Run rate (Q2 x 4) | ~$46B | ~$27B |
| H1 2026 operating result | -$0.7B | -$21.6B (-$9.3B Q1, -$12.3B Q2) |
| H1 operating cash flow | +$3.9B | Not reported |
| Cash at 6/30 | $131.7B | Not reported |
| Stock pay in the loss? | Yes ($2.2B) | Yes in Q2; often excluded in “adjusted” figures |
| Q3 run-rate growth | Not yet reported | 77% (~$50B at end of Sept) |
| Compute commitments | $518B (~80% binding) | ~$665B |
| Commitments / run rate | ~11x | ~25x |
| Main compute suppliers | Broadcom $161B, Google $111B, Amazon $110B, Microsoft $31B | Oracle, Microsoft, CoreWeave, AWS |
| Supplier credit | Mostly AA-rated hyperscalers | Oracle Baa2 negative; CoreWeave B+, bonds 15%+ |
| Valuation / run rate | $1.5T IPO talk = 33x | $1.4T = 52x |
Sources:
Anthropic and OpenAI sell frontier: the smartest models for now. “For now” is the problem. The best open-weight (open source AI) models trail by about four months. Many models come from China, where the Chinese government subsidizes energy, and the economy sustains lower development costs.
Chinese models can cost 1/20th of US models. DeepSeek V4 Pro charges $0.87 per million output tokens. Claude Opus 4.7 charges $25. GPT-5.5 charges $30. That is a 29x premium for a four-month head start.
Open-weight (open-source AI) models now handle 34% of enterprise AI tokens, up from 23% a year ago, and 69% of those users cite cost. So the frontier labs must keep running just to stay in place. Each new model buys four more months. Then the price floor drops again.
In 2026, big companies paid frontier prices. They were burning $100 bills in the parking lot. Uber gave 5,000 engineers Claude Code, and by April, its CTO said the entire 2026 AI budget was gone. In June, Uber capped spending at $1,500 a month per employee for each tool. Amazon used a leaderboard that ranked staff by token use, later finding employees gaming it on trivial tasks. Meta’s dashboard “Claudeonomics” was also taken down.
So where does the frontier actually pay off? Four areas stand out:
The real question is- do these really stand out? If the frontier gives only four months of advantage, why not wait months and code for free? What does a legal firm gain by creating a document handling system that must be rewritten every 120 days to stay ahead of competitors? And finally – wouldn’t effective employee management create far better projects than pushing expensive leading-edge projects onto burnt-out and unmotivated staff?
For corporate success – isn’t effective management far more important than having the coolest hot-rod in the garage?
Look at where the stress shows up. It is not in the products. It is in the bonds. On October 15, the cost to insure CoreWeave’s debt hit a record high. Traders were pricing better than 60% odds of default within five years. CoreWeave owes nearly $50 billion. Its 2032 bonds traded below 80 cents on the dollar. Moody’s rates Oracle just two steps above junk, with a negative outlook. Even the giants are borrowing. Hyperscalers are on track to issue about $175 billion in debt this year.
So is this all of AI, or just OpenAI? Follow the contracts. OpenAI signed a $300 billion compute deal with Oracle. CoreWeave and Microsoft both hold large OpenAI contracts. On October 8, OpenAI’s run rate came in at $50 billion, not the $70 billion reported earlier. Nvidia, Oracle and CoreWeave all fell that day. Anthropic’s big contracts point elsewhere. Its largest suppliers are Broadcom, Google, Amazon and Microsoft, and all of them have other income to fall back on. Most of the bond stress traces back to one weak link: OpenAI. They are just a bit over the front of their skis.
AI is not the only pressure on the bond market. Oil is another. Brent crude averaged $114 a barrel in September after attacks on Middle East oil sites. US inflation rose to 3.7% in September, mostly from gasoline. The Fed raised rates that same month, its first hike in over three years. Markets now expect another in December. Meanwhile, US debt has passed $40 trillion. Interest on it hit a record $1.05 trillion in fiscal 2026.
The US is not alone. In early October, the UK’s 30-year bond yield topped 6%, its highest since 1998. France’s 10-year neared 5%. Japan’s 30-year set a record. All the major economies are borrowing at the same time, and they compete for the same pool of savings. In normal times, a stumble at one AI company is a local problem. In this market, any bump in yields gets magnified. The weakest borrowers feel it first. Those are CoreWeave, Oracle and the private credit funds behind AI data centers.
QQQ is not a broad bet on tech. The top 10 holdings make up about half of the index. Many of the biggest are chipmakers that sell into the AI build-out. As of June, Nvidia was 7.5%, Micron 5.7% and AMD 3.9%. Companies like Oracle that are not QQQ still have chip orders that flow there.
South Korea’s KOSPI index has already fallen more than 40% from its June peak. The selloff centers on Samsung and SK Hynix, two of the world’s biggest memory makers, and it was driven by doubts about AI profits and hot US inflation. Taiwan sits in the same chain. Higher US yields also pull money out of other markets and into the dollar.
The risk is not that OpenAI will fail, but that rising bond yields, combined with a mild recession, will ripple through global debt markets and pose a systemic threat to global banking. The animal spirits may awaken and cause the very problem they fear most.
Is the AI crash industry-wide, or just OpenAI? The Anthropic S-1 gives us a roadmap. It shows that a frontier lab can grow fast, book real revenue and keep its commitments within reach. OpenAI appears to have taken on more than it can handle. One company with $665 billion in commitments, and suppliers who borrowed to meet them, is enough to make the whole market fragile. Add high oil, record US debt, and rising global yields and a small stumble can turn into a crash.
But a crash is not the end of AI. The technology works, and adoption is real. Low-cost, open-weight (open-source AI) models get better every four months. They put frontier-level tools in every business’s hands at a fraction of the price. That is how booms are built. The fiber laid by Global Crossing helped create the tech boom of the last two decades. It looks like AI is on this path. The survivors of the crash will build the boom based on new productivity, paid for by those who lose their investment.
Most of the risk centers on OpenAI. Its compute commitments are about 25 times its run rate, and its suppliers borrowed heavily to meet them. Anthropic’s commitments are about 11 times its run rate, mostly with large cloud companies, according to its S-1. When OpenAI’s revenue came in lower than reported, Nvidia, Oracle and CoreWeave all fell.
Not yet, but the warning signs are in the debt markets. CoreWeave’s default insurance hit a record in October 2026. South Korea’s KOSPI has fallen more than 40% from its June peak. The technology itself is still growing fast.
The 95% figure comes from a 2025 MIT study of corporate AI pilots. Many companies did pay frontier prices for weak results. Uber burned its 2026 AI budget in four months and then capped spending. The problem is wasteful spending, not the technology.
The technology won’t. The financing might. Companies that borrowed to build AI data centers face rising interest rates. CoreWeave’s bonds trade below 80 cents on the dollar. Some companies may fail, as fiber companies did in 2001, while the technology keeps spreading.
The hype is fading, but the use isn’t. Companies are moving from frontier models to cheaper open-weight models. Open-weight models now handle 34% of enterprise AI tokens, up from 23% a year ago. That is a price shift, not a loss of interest.
The US leads at the frontier. China leads in low-cost open-weight models such as DeepSeek and Qwen. Those models trail the frontier by about four months and cost a fraction of the price. In a price war, that lead may matter more than the frontier.
Worry about the debt, not the technology. High oil, rate hikes and record US interest costs make 2027 a fragile year for anyone who borrowed to build AI. For everyone else, cheaper AI tools are likely to keep getting better.
Managing dynamic contact databases, multi-device calendar sync, and corporate scheduling tools requires a rock-solid communication…
People regularly switch between smartphones, laptops, tablets, and desktop computers throughout the day. Whether working,…
Retaining top talent demands more than just a standard salary. Discover how growing companies are…
With growing businesses, the management of employees becomes quite complex. While it is possible for…
A starter set of business email templates for small teams: welcome, onboarding, thank-you and follow-up…
Compare the best mobile proxy providers of 2026 by speed, block rate and real price…
