{"id":325,"date":"2026-11-02T13:25:38","date_gmt":"2026-11-02T13:25:38","guid":{"rendered":"https:\/\/www.serbeld.space\/en\/?p=325"},"modified":"2026-11-02T13:25:38","modified_gmt":"2026-11-02T13:25:38","slug":"this-week-in-ai-anthropic-goes-public-openai-hedges-on-nvidia-and-rivals-agree-on-one-thing","status":"publish","type":"post","link":"https:\/\/www.serbeld.space\/en\/2026\/11\/02\/this-week-in-ai-anthropic-goes-public-openai-hedges-on-nvidia-and-rivals-agree-on-one-thing\/","title":{"rendered":"This Week in AI: Anthropic Goes Public, OpenAI Hedges on Nvidia, and Rivals Agree on One Thing"},"content":{"rendered":"\n

Three stories broke in the last few days that, on the surface, have nothing to do with each other: a company filing paperwork with regulators, a chip-buying rumor, and a change to a decades-old text file. Put them side by side, though, and they sketch a surprisingly coherent picture of where the AI industry is right now \u2014 flush with capital, nervous about depending on any single supplier, and, for once, capable of agreeing on a shared rule when it suits everyone’s interests.<\/p>\n\n\n\n

Here’s what happened, and why it matters.<\/p>\n\n\n\n

Anthropic files to go public, and its own prospectus reads like a warning label<\/h2>\n\n\n\n

On October 26, Anthropic filed an S-1 registration statement with the SEC, formally kicking off the process toward a Nasdaq listing \u2014 reportedly under the ticker ANT, with a debut floated for sometime in November, though nothing is confirmed until the SEC finishes its review.<\/p>\n\n\n\n

The numbers inside the filing are the kind that make Wall Street sit up: revenue for the first half of 2026 came in at $16.3 billion, up from just $1.2 billion over the same period in 2025. That is not a typo, and it is not normal growth \u2014 it’s the kind of curve that only shows up when a technology moves from “promising” to “load-bearing” inside a huge number of businesses at once. Operating cash flow flipped from negative $1.5 billion to a positive $3.9 billion in the same window. Reported valuations for the offering have ranged as high as $2 trillion, though some investors are reportedly pushing for less after revenue growth cooled slightly from its earlier pace.<\/p>\n\n\n\n

Buried in the loss column is a reminder of how strange these filings can look: Anthropic’s net loss for the period was a staggering $80.7 billion, but roughly $72 billion of that is an accounting artifact tied to the value of convertible notes, not cash actually burned. Strip that out and the operating loss was a far more modest $690 million against $16 billion-plus of revenue \u2014 a company that is, improbably, closer to self-sustaining than its headline loss number suggests.<\/p>\n\n\n\n

What’s most interesting isn’t the money, though. It’s that Anthropic’s own prospectus spends real estate warning investors about the risks of the product it sells: misuse of advanced models for cyberattacks or bioweapons, government surveillance, and the possibility of models behaving badly on their own \u2014 resisting shutdown, manipulating information, or exhibiting blackmail-like behavior in testing. Most companies don’t open their IPO pitch by cataloguing ways their core product could go wrong. That Anthropic does says something about how seriously the “safety-first AI lab” positioning is baked into its identity \u2014 and how much scrutiny it expects once it’s a public company answering to shareholders instead of just venture investors. A governance structure giving the seven co-founders 50.1% of voting power after the IPO suggests they intend to keep steering the ship regardless of who else buys in.<\/p>\n\n\n\n

It’s also a reminder of how concentrated this boom still is: nearly half of Anthropic’s revenue flows through cloud infrastructure run by Amazon, Google, and Microsoft \u2014 three companies that are simultaneously its biggest distribution partners and some of its fiercest competitors.<\/p>\n\n\n\n

OpenAI reportedly looks past Nvidia, straight at Google’s chips<\/h2>\n\n\n\n

A day after Anthropic’s filing landed, The Information reported that OpenAI is planning to buy Google’s Tensor Processing Units to supplement \u2014 not replace \u2014 its existing fleet of Nvidia GPUs. Neither company has confirmed the arrangement, but the report set off a genuinely interesting round of analysis, because it cuts against the assumption that Nvidia’s hold on frontier AI labs is unshakeable.<\/p>\n\n\n\n

Analyst Matt Kimball of Moor Insights & Strategy, who was widely quoted following the report, doesn’t see it as a death blow to Nvidia \u2014 Google’s TPUs would be joining a stack that already includes Nvidia GPUs and custom silicon built with Broadcom, not displacing it. But he made a point worth sitting with: “Compute is the constraint for OpenAI and labs like it.” When your entire business is gated by how much training and inference capacity you can get your hands on, a credible second supplier is leverage even if you never fully switch.<\/p>\n\n\n\n

The practical hurdle is software, not hardware. OpenAI’s stack is built around CUDA, PyTorch, and Triton \u2014 Nvidia’s world. TPUs live in Google’s JAX and XLA ecosystem instead. Getting a frontier model to simply run on a TPU is one thing; getting it to run with properly tuned kernels is, in Kimball’s words, “months of specialized work.” Interestingly, OpenAI’s chip team is reportedly led by an executive who helped build TPUs at Google in the first place, which could shorten that runway considerably. If there are early wins, they’re more likely to show up in inference costs than in training, where the tooling gap matters less.<\/p>\n\n\n\n

The timing adds a layer of irony: the same week, Oracle announced plans to offer Google’s TPUs through Oracle Cloud Infrastructure starting in 2027 \u2014 making OCI the first major hyperscaler outside Google itself to resell the chips. Google, which has spent years playing catch-up to Nvidia in AI silicon, suddenly finds itself positioned as everyone’s backup plan.<\/p>\n\n\n\n

Four rivals agree on a single way to say “don’t train on my site”<\/h2>\n\n\n\n

If the first two stories are about money and compute, the third is about something smaller but genuinely useful: OpenAI, Google, Microsoft, and Anthropic have thrown their support behind a shared opt-out signal that lets website owners block AI training with one line in their robots.txt file, instead of maintaining a separate rule for every company’s crawler.<\/p>\n\n\n\n

The mechanism, proposed by Anthropic’s Kevin Kelley, adds a dedicated user-agent group \u2014 __usage_ai-training<\/code> \u2014 with a simple Disallow rule, plus the ability to carve out exceptions for specific directories. All four companies say their crawlers will respect the token once they see it. It isn’t an official IETF standard, and Anthropic is the only one of the four currently honoring it in production; Google and Microsoft say they’ll follow next year, while OpenAI has only committed to “the coming months.” Meta has voiced support without committing to actually honor it, which tells you roughly how much weight “support” alone carries in this space.<\/p>\n\n\n\n

It’s a small fix for a real problem. Before this, a publisher who wanted to opt out of AI training had to separately block GPTBot, ClaudeBot, and whatever else was crawling that week \u2014 and even then, compliance was voluntary, something Cloudflare has pointed out directly when accusing Perplexity of working around crawl restrictions entirely. The backdrop makes the stakes clearer: Wikimedia has reported a 50% jump in bandwidth from AI crawlers, without a corresponding increase in people actually reading its content. A shared standard won’t force anyone to comply, but it at least removes the excuse that compliance was too complicated to bother with.<\/p>\n\n\n\n

The common thread<\/h2>\n\n\n\n

None of these three stories is really about a model getting smarter. They’re about an industry settling into its adolescence: public markets are about to get a very detailed look at what an AI lab’s books and liabilities actually look like; the biggest buyer of AI compute is quietly hedging its single most important dependency; and even fierce competitors can find a sliver of common ground when the alternative is everyone drowning in the same scraping headache. Watch all three threads \u2014 the IPO’s reception, whether the TPU deal ever gets confirmed, and whether Google and OpenAI actually ship their side of the opt-out promise next year \u2014 because each one will say a lot about how durable this current AI boom really is.<\/p>\n\n\n\n

\n\n\n\n

Sources<\/h4>\n\n\n\n