Anthropic hits 30 billion in annualized revenue as Meta goes complete token maxing. Welcome back to the AI Daily Brief — headlines, all the daily AI news you need in around 5 minutes. We kick off today with a big update in the competition between the labs, as Anthropic has announced that they've now reached 30 billion in ARR. It was actually tucked into a blog post about their new deal with Google and Broadcom, which we'll cover in just a minute, but that is a 3x increase since the end of last year and up 58% since the end of February. Now, according to the latest numbers that we have from OpenAI, that suggests that Anthropic has flipped them to have a higher annualized run rate. Although we've also heard in the past that they don't calculate things exactly the same way, and you better believe that if they haven't actually gone ahead of OpenAI in revenue, we will hear from OpenAI about it very soon.
Now, this all comes as the financials for both of these companies come under much greater scrutiny as they head towards an eventual IPO at the end of this year or the beginning of next. On Monday, the Wall Street Journal published a deep dive into OpenAI and Anthropic's numbers, sourced from financial disclosures around each company's recent fundraising. The key focus was on training costs, which are sky-high for both companies. OpenAI expects to spend around 30 billion on model training this year, which is triple what they spent last year. Anthropic's projected training costs are relatively more modest, but still almost triple to reach 28 billion by 2028. Now, while both training budgets are massive, it's notable that OpenAI is forecasting costs to go up on a completely different level than Anthropic. Because training costs are so high, both companies are providing an alternate counting of profitability that excludes them. Without training costs, both OpenAI and Anthropic are on track to eke out a small profit this year, with that profit accelerating moving forward.
Not everyone loves this financial engineering. Ram Maluwalia sums up the feeling of many investors when he writes, "OpenAI and Anthropic are incredibly profitable if you just strip out the training and inference costs. This business model is equivalent to running a passenger airline except you need to replace your jets every 6 months. Bizarre to have another definition of earnings simply because we don't like the costs." Now, in terms of top-line revenue, both firms expect to double revenue this year and are forecasting further doublings over the next few years. Notably, Anthropic's revenue is almost entirely from enterprise customers, and they forecast that to continue effectively indefinitely. OpenAI's revenue is more balanced than it used to be, but still skews towards the consumer. They do expect enterprise and consumer revenue to balance out over time. However, for the moment, this means OpenAI is spending money on inference for a ton of free users that Anthropic doesn't have to carry. OpenAI expects it to take until 2030 for them to turn cash flow positive, while Anthropic is forecasting a profit of the old, well-understood variety by 2028.
Now, the Wall Street Journal's analysis here is not particularly novel. We've had the rough contours of these financials from other sources already. What's more notable is that Wall Street is starting to analyze these companies as public market behemoths rather than growth-stage startups. The Journal had a very clear spin on the analysis, summed up by this closing line: "Both OpenAI and Anthropic will burn through a giant amount of cash in the coming years and are counting on their IPO investors to help buoy their businesses." TLDR: that is going to be the default narrative they fight against during their IPO and over the next few years. Still, for many people, the big story here is this massive new Anthropic number. Fleeting Bits points out, "Anthropic is growing at an annualized 9,700%. This is the fastest revenue growth at this scale in history. I don't know how to communicate the significance of Anthropic's growth rate at this scale without sounding hyperbolic." I asked Claude, and the best comparison that I could find was Nvidia, which grew at a 1,240% annualized rate during its best individual quarter growth ever, which was Q2 in fiscal year 2024. As John Arnold puts it, "Hard to believe that just 18 months ago Anthropic was broadly considered the odd man out of the AI race with an ambiguous business plan and no clear funding model. Not so anymore."
Now, on the back of soaring usage, Anthropic has signed a massive new compute partnership with Google and Broadcom. Anthropic announced on Monday that they've expanded their existing partnership to add multiple gigawatts of capacity set to come online from 2027. The Wall Street Journal added that the precise number is 3.5 gigawatts. Alongside the reveal that revenue had tripled to a 30 billion run rate, Anthropic also noted that enterprise spend specifically is skyrocketing. During their fundraising announcement in February, Anthropic boasted that 500 enterprise customers had annual spends above a million dollars. Less than 2 months later, that figure has doubled to 1,000 customers. Regarding the compute plans, Anthropic will build the majority of their new data centers in the US. The deal will expand Anthropic's commitment to deploying Google's TPUs, which are manufactured by Broadcom. Anthropic already began deploying TPUs in the fall and uses them exclusively for inference. Their training clusters are exclusively developed and operated by AWS, and that partnership remains ongoing. For Anthropic, the deal is obviously necessary. Their capacity constraints have become a huge problem this year, so they need pretty much every chip they can get their hands on. Yet, for Google and Broadcom, this is arguably even more important. Google set out to build a new business around external TPU sales last year. Many argued that they didn't have the sales or support staff to compete with Nvidia or AMD and would face a hard slog to set up a new business line. Yet now, in a single deal, Google has built a multi-billion dollar chip business around a solo customer. Broadcom, meanwhile, has guaranteed demand as long as Anthropic keeps growing.
Mohammad Hassan sums up, "The AI arms race just turned into a full-on power plant competition." Speaking of Google, after releasing their new open-source small model Gemma 4 last week, the company has wasted no time in productizing it. On Monday, they released an AI dictation app called Google AI Edge Eloquent. The product competes with things like WhisperFlow, allowing users to do live AI-assisted dictation on their phone. Edge Eloquent can filter out filler words, clean up phrasing to convey the intended message, and store custom jargon and keywords, much like the other AI dictation apps. The big twist is that everything is run completely locally on device. Users download the app with a packaged small language model and then can operate everything without an internet connection. Now, although Edge Eloquent probably isn't all that exciting, it does demonstrate a few interesting things about Google's Gemma 4 family. First, unlike previous small models, this doesn't seem to be a research project. It is a commercially viable model for certain use cases, and Google seems intent on building products around it. In addition, this could be the kind of local model Apple has been looking for to drive Siri, which is expected to use Gemini family models when it relaunches in the summer. Gemma 4 doesn't seem to be quite there yet for driving a full offline version of Siri, but you can see where Google is heading.
Now, aside from commercial applications, Gemma 4 has seen a hugely positive response from the developer community. Gemma 4 was downloaded 2 million times in its first week. In contrast, Gemma 3 received 6.7 million downloads over the past year, while Alibaba's Qwen 3.5 achieved 27 million downloads since its release in mid-February. Something that went a little under the radar is that the entire family of models, right down to the 2B version, have strong agentic performance that could push the frontier for mobile agents. Philip Schmidt, a developer experience lead at DeepMind, showed the model can query Wikipedia using agent skills while running on an iPhone. Obviously still very early innings, but it feels like Gemma 4 could lead to a breakout moment for local models, especially once the Open Claw folks start tinkering with it.
Over in Meta land, the company is preparing to release their new model and plans to offer an open-source version in the future. Axios published new information about the model release on Monday, citing sources familiar with the views of AI CEO Alexander Wang. They wrote that Meta wants to keep some part of the model proprietary during the initial release to ensure it doesn't introduce new levels of safety risk, and this reporting contradicts prior speculation that Meta would abandon their commitment to open-source models as part of this new release. Axios added that an open-source model aligns with how Wang sees Meta's position in the AI race. Wang reportedly views Meta as a democratizing force that can ensure there is a US-trained option for open-source developers. Sources suggest that Wang believes that OpenAI and Anthropic are increasingly focused on developing AI systems for governments and the enterprise, while Meta is focused on the consumer. Writes Axios, "Meta wants its models distributed as widely and as broadly as possible around the world." Now, this is the first news we've had on the forthcoming model, code-named Avocado, in several weeks. In early March, The New York Times reported that Avocado had been delayed and couldn't match Gemini 3 on benchmarks. Talk of safety concerns could imply that model performance has improved with another month of post-training. Still, sources say that Meta knows that its models won't be competitive across the board, but believes they will have certain strengths that drive consumer appeal.
Meanwhile, while Meta's own model is getting close to release, Meta engineers are still using Claude — a whole lot of Claude. The Information reports that Meta employees have set up an internal leaderboard to see who is churning through the most tokens. The leaderboard is dubbed Claudonomics and aggregates the top 250 token users among Meta's 85,000 employees. Top-ranking token users can earn the rank of session immortal or token legend. The Information argued this is a new type of conspicuous consumption in Silicon Valley known as token maxing. The thought is that token consumption is a good proxy for AI-enhanced productivity, so engineers want to climb the leaderboard. Now, the flaw in thinking is immediately obvious, with The Information also reporting that some at Meta are running large numbers of agents in parallel with the goal to rip through as many tokens as possible, not necessarily be as productive as possible.
And while token maxing could be the new version of judging engineers by counting how many lines of code they write, the culture is being driven from the top. Last month, Nvidia CEO Jensen Huang said he would be deeply alarmed if an engineer on a $500,000 salary wasn't using $250,000 worth of tokens annually. That's also the view at Meta, with CTO Andrew Bosworth boasting in February that one of his top engineers is spending the equivalent of his salary on tokens to generate a 10x efficiency boost. Bosworth commented, "This is easy money. Keep doing it. No limit." Now, there is a lot of chatter on this, with many feeling like Joe Weisenthal, who writes, "How does measuring productivity by total token consumption make any sense at all?" Comparing it to Chairman Mao requiring peasants to smelt steel in their backyards during the Great Leap Forward, which of course led to tons of useless low-grade steel. Joe continues, "Real backyard steel furnaces vibe, in my opinion." Interestingly, Meta Critic Capital on Twitter makes a different China comparison to argue why this actually makes sense. He wrote, "Very early in its development, the party would set GDP growth goals for each province in China. As you know, you can do lots of silly things to boost GDP growth, and party officials certainly did. But the opportunities for China's development were so vast that simply putting a GDP growth target was enough. It took decades for Goodhart's law to catch up with them. Same goes for tokens. Meta is spending 90 million tokens per developer per day. At Opus 4.6 rates, Meta would be spending in the zip code of 4 to 5 billion dollars per year. I think all but five corporations on Earth can spend that much on AI. It's a massive feat of engineering from wall to wall to be capable of spending that many tokens." TLDR: the cost of token maxing is small because token maxing is extremely hard. You can safely expect that over the next 18 months, 98% of corporations would be better off token maxing. Interesting thoughts there, but for now that is going to do it for today's headlines. Next up, the main episode.
