Analysis
Anthropic is reportedly negotiating to pay roughly $6 billion for Decart, an Israeli startup whose core product isn't a chatbot or a coding agent -- it's software that makes existing GPUs run more efficiently, according to Bloomberg. That would be Anthropic's largest acquisition to date, and on its own it reads as one more entry in a summer full of AI M&A headlines. Looked at alongside two other things that happened this same week, it reads as something more specific: a market where the thing everyone is actually fighting over is compute access, and the model layer is downstream of that fight.
Start with why Decart is worth that much to begin with. Pulse covered Decart's earlier raise in June, when it priced at roughly $4 billion on the strength of real-time video generation and 'world model' research. The reported acquisition price is a roughly 50% premium over that mark, set barely ten weeks earlier -- and the premium isn't for the video product.
Reporting on the talks makes clear Anthropic wants Decart's chip-efficiency software specifically, because it lets Anthropic's existing GPU fleet absorb more inference and training demand without buying a single additional chip. In a year where GPU allocation is still the scarcest input to how fast any lab can grow revenue, buying your way to more effective capacity is cheaper than buying more literal hardware.
“The reported acquisition price is a roughly 50% premium over that mark, set barely ten weeks earlier -- and the premium isn't for the video product.”
## Three deals, one shared logic That's deal one. Deal two is IBM's newly announced enterprise partnership with OpenAI, giving IBM's consulting and enterprise-software customers direct access to GPT-5.6 inside IBM's own stack. On the surface it's a product integration. Functionally, it's IBM renting distribution to OpenAI in exchange for OpenAI renting IBM's decades of enterprise relationships -- neither company has to build what the other already has. Deal three is older but still doing work in the market: Tencent president Martin Lau's comment on the company's own Q2 earnings call that Tencent could generate a 'decent return in an immediate timeframe' by renting out its roughly $53 billion in AI infrastructure spend instead of using it to run first-party models. That's a company that just tripled its AI capex year-over-year saying, out loud, that the chips themselves might be worth more than what it's currently doing with them.
None of these three companies coordinated with each other. But line them up and the shared logic is hard to miss: Anthropic buying efficiency software to stretch its GPU fleet further, IBM trading distribution for a model it didn't have to train, and Tencent openly weighing whether to become a compute landlord instead of a model builder. In each case, the scarce asset being traded is usable compute -- not IP, not brand, not even revenue growth. Compare that to how AI M&A got priced a year ago, when acquirers were mostly buying research teams and model weights. The unit economics have shifted toward whoever can squeeze the most usable inference out of a fixed GPU budget.
## The numbers behind the thesis Anthropic is sitting on more than $47 billion in annualized revenue and is reportedly targeting an IPO valuation that could approach $2 trillion -- a company with that kind of growth doesn't need Decart's balance sheet or its headcount. It needs the thing Decart's software actually does: more inference per dollar of GPU spend, at a moment when Anthropic's own compute bill is almost certainly its largest single cost line.
Meanwhile SpaceXAI's all-stock acquisition of Cursor-maker Anysphere earlier this year followed a similar shape -- pairing SpaceX's compute-heavy infrastructure with a coding-agent business that needed more inference capacity than it could otherwise afford on its own. Different sectors, same trade.
For founders, the read is straightforward: if your startup's real IP is a layer of software that makes someone else's GPUs go further -- inference optimization, scheduling, quantization, anything adjacent to Decart's category -- you are currently more interesting to a strategic acquirer than a startup with a slightly better chatbot. For GPs marking portfolio companies in this category, the diligence question isn't whether the technology works in a benchmark; it's whether the efficiency gains hold up on someone else's production fleet at scale, because that's the only claim these acquisitions are actually paying for.
The bear case here is that not every AI deal fits this frame, and forcing all of them into a compute-access story risks overreading three data points into a movement. Lovable's recent mega-round was about product velocity and revenue multiple, not GPU access. Cognition's reported valuation talks are about coding-agent market share, not chip efficiency.
The compute-access thesis explains a specific, narrower category of deal -- acquisitions and partnerships involving labs that are themselves compute-constrained -- and it would be a mistake to apply it to every large funding round this quarter.
Watch whether the Decart talks actually close, and at what final price -- a deal that slips below $6 billion or falls apart entirely would say the 50% premium was more negotiating theater than real conviction. Watch too whether other labs follow with their own efficiency-software acquisitions rather than pure model-team buys; if Google, Meta or xAI announce something in Decart's category within the next quarter, this stops being three data points and starts being the actual shape of 2026 AI M&A.
