Six Anthropic investors told the FT the company is on track for $100bn to $120bn of annualised revenue by the end of 2026, which is roughly the growth rate it was running at when it disclosed a $47bn run rate in May, and would put it somewhere around $70bn now. On that basis they are floating an IPO at $2tn, and one of them helpfully observed that a company growing 800% a year should trade at 30 times revenue, which gets you to $3tn. These are people talking their book, and the company itself has not named a number. What I would like to see is actual GAAP revenue (most reporting suggests the run-rate figure is what end customers pay cloud providers, not what Anthropic itself books), plus gross margins and the real cost structure. The deeper problem with valuing any of these companies is that nobody knows where supply and demand settle, how much pricing power the labs keep, how long the frontier race stays this expensive, or where we are in the cycle, if it even is a cycle. If you bought Amazon in August 1999 you were completely right about the future, down 90% two years later, and not back in profit until 2007. There is nothing to anchor the analysis, which is why SpaceX trades the way it does.
Meta's reset continues. Its new models are close to the frontier and, more usefully, sit on the right part of the price-performance curve, and now it has gone back to releasing them openly: Muse Glimmer, a 30-billion-parameter model under Apache 2.0, built to run agents on a single consumer GPU, with the weights of the larger Muse Spark 1.2 promised next. You no longer have to go to China for a good open model. Meta's strategy was always that models should be commodity infrastructure sold at cost. It has no cloud business to sell them at a margin, but its own products depend entirely on what AI can do, so it wants the models cheap and the value captured on top, in social and ads. This time there is a second layer, though. Releasing an open frontier-class model is a statement about risk and politics. Anthropic has argued that the dangers of this technology, even ordinary security dangers before you get anywhere near AGI, mean access to the frontier should be tightly controlled. That may well be a sincere position. It is also, unavoidably, an argument for fewer competitors and a regulatory moat. Meta is taking the opposite line, that safety comes from giving everyone the same tools. Which is, of course, exactly the attitude it took to social media.
On to the financing. Two weeks ago the Journal said Nvidia would guarantee $250bn of borrowing for OpenAI's Ohio data centre; that has since been cut to roughly half. Meanwhile Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, targeting more than $500bn of outside capital. The simple version is that Nvidia wants GPUs to be a tradable asset you can borrow against: if you build an AI cloud on debt and go bust, someone else picks up the chips, so the lender's risk is lower. That is clever, and it rests on two load-bearing assumptions. First, that GPU resale prices are predictable enough over the life of a loan to work as collateral. These are not obviously long-lived assets. Second, that defaults won't be correlated. The most likely reason anyone fails to service a GPU loan is that compute prices have collapsed, in which case nobody else can service theirs either, and nobody wants the collateral. The fibre boom is a poor analogy in most respects (fibre was built ahead of demand, this is being built behind it), but it would not have been a great idea to lend to fibre carriers against the fibre. And underneath all of it, the cash flow that backs Nvidia's own guarantees is exposed to the same correlation.
Anthropic made some noise implementing the EU's new requirement that AI text carry a machine-readable mark. The method is a version of the SynthID approach: at the points in a sentence where several words would do equally well, the model's choices are nudged into a statistically detectable pattern. That means no watermark if you only used the model to fix typos, because it never got to choose, and a watermark that thins out on factual text and code where there is one right answer. Heavy rewriting degrades it, and in principle you could build a tool to find those interchangeable words and re-roll them. I don't think text can be watermarked robustly; images are a different matter, there is far more signal to hide things in. Anthropic is applying it worldwide, not just in Europe, because it has no clean way to scope it by region.
Then a story that will be the first of many, and it happened down the road from me. A developer in Melbourne asked his agent, OpenClaw running on a Claude model, to get him into a full Pilates class. The agent found that the gym's booking API checked who you were when you made a booking but not when you cancelled one, so it cancelled a stranger's reservation, moved its user up the list, and then admitted it could not put the stranger back. Two problems in one anecdote. AI collapses the cost of hacking, so a great many systems that were never worth attacking are now in play. And an agent will do what you told it, often successfully, which may not be what you thought you told it. At the other end of the scale, more than forty bitcoin and crypto firms published an open letter complaining that safety restrictions on frontier models leave defenders with weaker tools than attackers, who are not asking anyone's permission. There is no clean answer here. It is trade-offs all the way down.
America and China. Having had the FCC ban imports of new foreign humanoid and legged robots last month, the White House signed tariffs of 100% on heavier and thermal-imaging drones and 25% on the rest. Part security, part protectionism, as all of this is. Illustrating the security half, the Royal Navy found that cameras on its new uncrewed surface vessels were sending heartbeat pings to a server in China, which may be nothing more than default settings on an off-the-shelf component and is still exactly the problem. And illustrating the unintended consequences, Reuters reports that Samsung and SK Hynix have been testing Chinese etching tools for their Chinese fabs as a hedge against losing access to American equipment. Jensen Huang's argument about export controls applies here too: you slow China down, and you build its domestic industry at the same time.
Narrative violation: Reuters says Silver Lake is in talks to take Workday private, at a market value of around $43bn before the leak. Workday is everyone's favourite example of legacy SaaS waiting to be eaten by AI. Maybe this is just the oldest lesson in markets: there are no good or bad companies, only wrong prices, and if the price is too low and you think you can fix it, you buy.
Three more. AT&T says open models handle about a quarter of its AI workload and it expects that to reach 70 to 80%; it is running about 45 billion tokens a day, a metric of uncertain meaning, and routing work to whichever model is good enough at the right price. That is what real enterprise deployment looks like: most jobs don't need the frontier, and nobody wants to depend on a single supplier, whether that supplier is Dario or Donald. The Journal points out what every analyst already knows, that Alphabet's record $112bn quarterly profit was mostly paper gains on its stakes in SpaceX and Anthropic, so the EPS isn't the earnings, and with the leases, guarantees and SPVs piling up, the capex isn't the capex and the debt isn't the debt either. And Indian workers are being paid around 250 rupees an hour, about $2.60, to wear head cameras while they work, producing the first-person footage that teaches robots to do those same jobs.
Data point: Sensor Tower has ChatGPT's share of AI assistant usage at 46%, under half for the first time, with Gemini at 28% and Claude at 10%. Distribution through Android and Search is doing for Gemini what nothing else could.