TT Tech Logo
All articles & videos
Article · English#AI#AI bubble#architecture#engineering leadership

The AI bubble will burst. Will your codebase survive?

AI money looks like the dot-com bubble. What a crash, a price rise or a lost vendor means for teams that build with agents, and what to set up now.

Adil Ben Moussa
Adil Ben Moussa
03 November 2026· 5 min
The AI bubble will burst. Will your codebase survive?

At the end of October, Anthropic filed to go public, and I read its . Anthropic makes the model I work with every day, so this is not a comfortable read. I think the AI bubble will burst, and I hope I am wrong. My team writes no code by hand any more: agents write it, we review it and decide what ships. I enjoy it more than writing code myself. We have superpowers now, and that is why the numbers worry me.

The money

Its revenue went from $4.6 billion in 2025 to $33.2 billion in the first nine months of 2026, with a net loss of $104.0 billion. Most of that loss, $94.3 billion, is a non-cash charge, mainly because its convertible notes are worth more now the company is valued higher. Operations brought in $6.6 billion of cash.

What worries me is the other column. It has committed about $895.4 billion of spending, "substantially all" of it compute and infrastructure, and only $213.8 billion of its obligations can be terminated or depend on conditions. Against that it holds $95.6 billion in cash and $36.3 billion in short-term investments. Some compute suppliers "have strategic and investment relationships with us and are our customers".

For 2026 Amazon expects about $220 billion of capital spending, Microsoft about $170 billion for the calendar year and Meta $135 to $145 billion. Amazon's free cash flow is now an outflow of $42.4 billion over twelve months.

Money that goes round

Amazon put $18 billion into Anthropic and $50 billion into OpenAI, and OpenAI committed $138 billion to AWS. Anthropic committed at least $110 billion to AWS and $121.3 billion to Google, which invested $10 billion in it in May 2026 and can add up to $20 billion more. OpenAI contracted $250 billion more of Azure from Microsoft, which owns about 24% of it. Nvidia guarantees up to $105 billion tied to data-centre leases for an OpenAI affiliate, and invested $2 billion in CoreWeave, which runs only Nvidia GPUs. AMD gave OpenAI a warrant for up to 160 million shares alongside a 6-gigawatt GPU deal.

Is this only a bubble story? No. The revenue is real: Google Cloud grew 103% year on year, Azure 48%. Real usage and a bubble can exist together. 2000 had both.

What the dot-com bubble teaches

When the dot-com bubble burst, the Nasdaq fell 77.9% from March 2000 to October 2002. The overbuild ran on a myth that internet traffic doubled every 100 days. Real traffic doubled about once a year, far too slow to fill the capacity being laid. Amazon's shares were down "more than 80%" in a year, and Bezos opened his letter with one word: "Ouch."

A crash wipes out valuations, not the technology.

Even the top plans may be subsidised

In July 2025 Anthropic announced weekly limits after "one user consumed tens of thousands in model usage on a $200 plan". In January 2025 Sam Altman wrote that OpenAI was losing money on its Pro subscriptions. GitHub was the clearest: it "has absorbed much of the escalating inference cost", and since 1 June 2026 every plan bills on usage.

Even the highest subscriptions may be subsidised. What happens to your team when that stops?

Two scenarios for your team

Prices go up, limits get tighter. The agent that ran all afternoon now runs for an hour, and someone chooses which tasks still get it.

A vendor drops out or changes its terms. Your rules live in one tool's config, your workflows in one tool's commands. You switch, and half your setup stays behind.

What helps

  • Rules in . An open format that , , Copilot and about twenty other tools read, and reads it when there is no CLAUDE.md.
  • Checks that run without a model. Lint, typecheck, tests. They still say no when the agent is pricier or gone.
  • Code the team can explain. Code nobody understands is . It stays cheap only while agents stay cheap.
  • An open-weight fallback you have tried once. OpenAI's gpt-oss, Mistral's Devstral Small 2 and Google's Gemma 4 are all Apache 2.0: free to download and run on your own machine, no subscription. You pay for the hardware, not the . They are weaker; learn where before you need them.
  • Let juniors build the skill, not borrow it. I wrote code by hand for more than twenty years before agents came, so I can judge what they write. A junior who only prompts has nothing to fall back on. When your team has time, let juniors write some features by hand and explain every diff they merge.

On Monday

Copy your agent rules into an AGENTS.md at the root. Then give one real ticket to an open-weight model and write down where it fell short.

Sources

  1. Anthropic, Form S-1 Registration Statement, SEC, October 2026
  2. Anthropic, Form S-1/A, Amendment No. 1, SEC, November 2026
  3. Amazon, Amazon.com Announces Third Quarter Results, Form 8-K, October 2026
  4. Microsoft, Fiscal Year 2027 First Quarter Earnings Conference Call, October 2026
  5. Meta, Meta Reports Third Quarter 2026 Results, Form 8-K, October 2026
  6. Amazon, Form 10-Q, quarter ended 30 September 2026, SEC
  7. Microsoft, The next chapter of the Microsoft–OpenAI partnership, October 2025
  8. Microsoft, Form 10-Q, quarter ended 30 September 2026, SEC
  9. Nvidia, Form 10-Q, quarter ended 26 July 2026, SEC
  10. CoreWeave, Form 10-Q, quarter ended 30 June 2026, SEC
  11. AMD, AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs, Form 8-K, October 2025
  12. Alphabet, Alphabet Announces Third Quarter 2026 Results, Form 8-K, October 2026
  13. Microsoft, Microsoft Announces First Quarter Results, Form 8-K, October 2026
  14. Federal Reserve Bank of St. Louis, NASDAQ Composite Index (NASDAQCOM), FRED
  15. Andrew Odlyzko, Internet traffic growth: Sources and implications, University of Minnesota, 2003
  16. Jeff Bezos, 2000 Letter to Shareholders, Amazon, 2001 (reproduced by The Seattle Times)
  17. Anthropic, Thread on X about weekly limits for Claude Pro and Max, July 2025
  18. Fortune, Sam Altman says OpenAI is losing money on Pro subscriptions, January 2025 (press report of Altman's post)
  19. GitHub, GitHub Copilot is moving to usage-based billing, April 2026
  20. Agentic AI Foundation, AGENTS.md
  21. Anthropic, Claude Code CHANGELOG, version 2.1.277
  22. OpenAI, gpt-oss-120b, Hugging Face, 2025
  23. Mistral AI, Devstral-Small-2-24B-Instruct-2512, Hugging Face, 2025
  24. Google, gemma-4-31B-it, Hugging Face, 2026