Bloomberg reports Anthropic is in talks to acquire Decart — not for its AI-generated video products, but for the GPU-optimization stack underneath them, and the inference team that built it.
Reported Deal — Key Markers
2023 — Founding
Decart founded in Israel, headquartered in San Francisco; builds GPU-optimization inference stack alongside real-time world models Oasis and Lucy.
October 2024 — Oasis Launch
Oasis, a real-time AI-generated Minecraft-style world model, draws roughly one million users within days of release.
May 2026 — $300M Raise
Decart raises $300M at roughly $4B valuation, backed by Nvidia; GPU-optimization stack licensed to cloud providers and AI labs.
August 13, 2026 — Bloomberg Reports Talks
Anthropic reported in talks to acquire Decart at approximately $6B — roughly 50% above the May mark. Deal unfinalized; Bloomberg explicitly notes talks could fall through.
Bloomberg reports that Anthropic is in talks to acquire Decart, the Israeli-founded, San Francisco-based AI startup, at a price of approximately $6 billion. That figure is reported and unfinalized — Bloomberg is explicit that the deal could fall through — and it should not be treated as settled. If it does close near that number, it would represent Anthropic’s largest known acquisition, and would mark Decart up roughly 50 percent above the approximately $4 billion valuation the company carried after its $300 million raise in May 2026, a round backed by Nvidia.
Decart is easy to misread on first pass. Its public face is two real-time world-model products: Oasis, an AI-generated interactive environment frequently described as an AI-built Minecraft that drew roughly a million users within days of its October 2024 launch, and Lucy (also marketed as Mirage), a system that edits live video from text prompts. These products attract attention. They are not, by Bloomberg’s framing, what Anthropic is primarily after. What Anthropic appears to want is the layer underneath: Decart’s GPU-optimization and inference stack — sometimes called DOS internally — which the company has licensed to cloud providers and AI labs as a revenue-generating product in its own right. Bloomberg’s reported rationale is specific: the technology would help Anthropic’s infrastructure absorb more demand, and Decart’s team would join Anthropic’s inference and performanceorganization.
Anthropic has no video product. It is not, as far as Bloomberg’s reporting indicates, acquiring Decart to build one. The world-model products appear incidental to the transaction’s logic. That reading — that Anthropic is buying efficiency, not models — is analysis supported by Bloomberg’s details, but Anthropic has not stated its intent publicly, and it would be wrong to conclude the company intends to discard Oasis or Lucy. What Bloomberg’s framing does establish clearly is where the reported value sits: in the inference stack and the team that built it.
The key insight: Anthropic is reportedly paying a model-company price for what is, in structural terms, an infrastructure team and a cost-reduction stack. The video products are the wrapper. The inference optimization engine — and the people who built it — is the reported acquisition target. This is the efficiency lever pulled from the demand side of a supply-constrained industry.
The Structural Read
The week’s throughline in AI infrastructure has been compute as the binding constraint — the supply wall in memory, power, and fabrication, and the financing machinery being built around it, from data-center lease structures to compute treated as an asset class. This reported deal is the demand-side answer to the same pressure. If you cannot acquire enough chips, and the chips you can acquire are expensive to run, the remaining lever is to make each chip do more work per dollar. Buying an inference-optimization team is a direct pull on that lever.
Decart’s own marketing claims video generation costs falling from thousands of dollars per hour to under a quarter per hour. Those figures are the company’s own, unverified by independent parties, and belong in quotation marks rather than in a conclusion. But the direction of the claim is consistent with what Anthropic would need: not just cheaper inference in the abstract, but the kind of step-change efficiency that moves unit economics materially before a public offering. Bloomberg frames the timing in exactly these terms — this is reported as happening as Anthropic prepares a closely watched IPO, itself reported preparation rather than a certainty. Lowering cost-per-token improves the unit economics you are about to show public markets. Vertically integrating the inference stack converts a variable cost you license from others into capability you own.
One important caveat on the efficiency logic: cheaper inference does not automatically mean lower total spending. Jevons’ paradox applies — when running a model becomes cheaper, usage tends to expand to consume the savings and then some. No single acquisition dissolves a supply wall built from memory, power, and fabrication constraints. The efficiency lever is real and consequential; it is not a resolution.
The Fifth AI Bottleneck — Business Engineer
“The competitive frontier for AI labs is moving from capability toward cost. Owning the efficiency layer stops being operational and becomes strategic — because the lab that can run its models cheapest sets the price floor for everyone else.”
The shape of this reported deal also fits a pattern worth naming. The industry keeps paying model-company valuations — the kind of multiples attached to frontier AI products — for what are, in substance, acqui-hires of infrastructure teams. The flashy product (Oasis, in this case) provides the public narrative and the price justification. The actual transfer is the people and the stack. The same dynamic appeared in how the industry priced infrastructure bets relative to Nvidia’s moat — the productstory and the infrastructure reality diverge, and the infrastructure reality is what survives the integration.
Three Implications
THE PRE-IPO BALANCE-SHEET LOGIC
If this deal closes, Anthropic arrives at public markets having vertically integrated a piece of its inference stack rather than renting it. That changes how investors read the coststructure — owned efficiency capability reads differently than licensed efficiency. The timing is not incidental; it is the balance-sheet argument for doing this now rather than after the offering. The IPO is itself reported preparation, not a certainty, but the logic holds regardless of precise timing.
BUYING THE TEAM AND STACK, NOT THE PRODUCT
The pattern this deal represents is more important than the deal itself: frontier labs are willing to pay model-company prices for infrastructure teams embedded inside product-facing companies. The product (Oasis, Lucy) provides the narrative and the valuation anchor. The actual acquisition target is the people and the low-level optimization work they have already done. Any lab evaluating M&A in the current environment should read its targets at this level — what is the product, and what is the stack underneath it.
THE EFFICIENCY COMPETITION IS NOW EXPLICIT
When Anthropic is reportedly willing to spend approximately $6 billion on inference optimization ahead of an IPO, it signals that the cost-per-token race has moved from an engineering priority to a strategic one visible to boards and investors. The labs that can run frontier models most cheaply set the competitive floor. Efficiency is no longer a secondary concern that gets funded after capability research — it is, by this reported valuation, worth as much as a model company.
The supply wall in chips, memory, and power is the bottleneck everyone sees. The Fifth Bottleneck is the efficiency layer — the inference and optimization stack that determines how much useful work each chip actually does. As the cost of compute becomes the variable that decides who survives, labs are no longer competing only on model quality. They are competing on cost-per-token at scale. This framework maps where that competition is happening and who holds structural advantage across the AI stack.
Bloomberg’s reported talks are unfinalized, the approximately $6 billion price is not confirmed, and Decart’s efficiency claims are its own marketing — hold all three carefully. What survives the hedges is the signal: Anthropic is reportedly prepared to pay a frontier-model valuation for an inference-optimization team and the GPU stack it built, because the cost of running models at scale has become a strategic variable, not an engineering footnote. Whether this specific deal closes is, in a real sense, secondary. The fact that it is being discussed at this price, at this moment, tells you where the competition among frontier labs is moving — from how capable the model is, to how cheaply you can run it.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
