When Anthropic came forward on April 6, 2026, to announce their new series of agreements with Google and Broadcom, they spoke about securing "multiple gigawatts" of compute capacity. This new infrastructure, which is being built on next-generation TPUs, is set to start coming online in 2027. At the same time, a regulatory filing from Broadcom clarified that the actual scale is closer to 3.5 gigawatts. At this sort of volume, Anthropic is really moving beyond the typical startup label; they are positioning themselves as a serious infrastructure force that is operating at the scale of a decent-sized regional power grid.
The company also shared that their annualized revenue run-rate has now climbed past the $30 billion mark. To put that in some context, they were sitting at about $9 billion when 2025 came to a close, and they were around $14 billion during that Series G round back in February. What’s even more striking is how quickly their high-spending clients are growing—the number of business customers who are paying more than $1 million a year has doubled from 500 to more than 1,000 in less than two months. In the world of enterprise software, you just don't see a growth trajectory like this very often.
The industrial logic beneath the surface
A lot of the initial reporting has framed this move as if it’s just a typical case of a supplier and a customer expanding their relationship, but if you look at it that way, you might end up missing the deeper strategic logic that is actually at work here.
What Anthropic, Google, and Broadcom are really doing is taking steps to reserve capacity early in a market where the primary bottleneck is no longer just about having enough chips. The real constraint is shifting toward things like deliverable megawatts of power, how the racks are integrated, and the supply of memory. Anthropic has even put out policy language suggesting that a single frontier model will eventually require gigawatts of power on its own, and they seem to expect that the U.S. AI sector will need at least 50 GW over the next few years. That lines up with what the International Energy Agency is seeing, too—they are forecasting that power demand from data centers will grow more than twofold by 2030.
The long-term arrangement that Broadcom has in place with Google is set to run through 2031, covering the development of custom TPUs as well as various networking components. There is one particularly notable line in the regulatory filing which mentions that Anthropic’s access to that capacity is "dependent on its continued commercial success." Even though that sounds like standard corporate boilerplate, it really does summarize the entire strategic bet that is being made here.
The question of revenue and its underlying quality
The fact that their run-rate jumped from $9 billion to $30 billion in just three months does point toward a genuine level of demand for things like agentic coding and automating enterprise workflows. And because the number of million-dollar customers has doubled, it suggests that this growth is being spread across the wider market, rather than being tied up in just a few experimental contracts.
But it’s important to remember that a run-rate is mostly a measure of velocity; it doesn't really tell you much about how durable the margins are or how diverse the customer base is. It doesn’t reveal how deep the discounts might be, or whether the current adoption is mostly being driven by temporary experimental budgets that companies have set aside. While it’s true that Anthropic’s run-rate technically exceeds the last figure we saw from OpenAI, the comparison is never going to be perfect since they have different ways of reporting their revenue. The bottom line is that both companies are still unprofitable, and the current thinking is that they will stay that way through 2029.
So even though the demand has been proven, the economic quality of that business is still something that remains to be seen.
Looking at the winners beyond the initial stock move
When the news first broke, Broadcom’s shares rose by about 3% while Alphabet’s stock remained mostly flat. This is a reaction that I think might be misidentifying where the most significant structural value is actually being created.
Broadcom is deeply involved in Google’s TPU roadmap all the way through 2031, which allows them to capture the economics of both the custom silicon and the networking side of things. It is certainly a clean way for investors to play the trend toward heterogeneous compute, but that sort of optionality is something the market has already recognized and mostly priced into the stock.
On the other hand, what Google has gained is harder to measure, but it might actually be deeper. They have been trying to prove that their TPUs can be competitive with Nvidia, that Google Cloud is capable of hosting massive frontier workloads for companies other than themselves, and that the infrastructure they built for internal use has real gravity for outside tenants. Having a company like Anthropic as a primary tenant does more to validate these claims than any internal benchmark ever could. This represents a significant boost in credibility for Google Cloud, even if the market largely chose to ignore it.
The risks that are hidden within the physical grid
The bear case here is really a structural one. If the leaders in the AI space keep reserving massive amounts of capacity, but the physical work of delivering that capacity slows down—or if the price of tokens starts falling faster than the cost of running the systems—the industry could end up facing a huge overcapacity problem by 2027. They might find that they have reserved more computing power than they can ever actually monetize in a way that is profitable.
When you’re talking about a 3.5 GW plan, you’re talking about land, getting connected to the power grid, and securing transformers—and these are all things that are currently facing very real supply constraints. The memory markets are still tight, and there is a growing level of local opposition to building new data centers in many areas. The reality is that capital does not just convert cleanly into productive AI tokens; that path is a long one, and it is becoming increasingly political.
Thinking about the AI industry's core
If you look at the big picture, Anthropic is trying to pre-buy a seat at what is becoming the industrial core of AI, while Google is busy trying to prove that their TPUs are that core. Meanwhile, Broadcom is basically sitting there and tolling both sides of the bridge. These are all rational bets for the companies involved, but none of them are completely de-risked. And so far, the market has not yet found a way to price the difference between simply reserving capacity and actually being able to generate a return from it.
not investment advice
Sources: https://www.anthropic.com/news/google-broadcom-partnership-compute