Anthropic Walked From Its $6B Decart Bid. Don't Fix Your Rate.

Your vendor bid $6 billion to make the thing you are about to buy cheaper to produce, then walked away from it. The bid is dead. The reason for it is not — and the clause that captures a falling cost curve is still not in anyone's standard paper.

Update — September 8, 2026: Anthropic walked. Bloomberg reported on September 8 that the company completed due diligence on Decart and then decided against the acquisition, at a price it puts at roughly $6 billion; representatives for both companies declined to comment, and one source said the two may still find other ways to collaborate. Nothing was ever signed. The rest of this piece stands, because the thing that produced the bid did not leave with it: the inference-margin problem, the multi-gigawatt capacity commitments from 2027, the confidential IPO filing and the documented tokenizer change are all unchanged. If you were about to fix a per-token rate for two years, the answer is still no.

Anthropic spent most of a month trying to buy Decart, an Israeli AI-infrastructure company. Bloomberg broke the story on August 13 at a $6 billion price, describing early-stage talks that could fall apart. Five days later Calcalist reported that advanced drafts had been exchanged, put the figure at roughly $7 billion paid mostly in Anthropic shares, and said the deal could be signed as soon as the following month. It never was, and the number that survives is Bloomberg's — a walk-away after diligence at ~$6 billion. If you are negotiating a multi-year Claude commitment this quarter, that history is still not trivia. A bid is a dated, public statement about where your vendor's unit costs are heading, and it does not stop being one because the buyer changed its mind.

---

What Anthropic Was Buying for $6 Billion

Anthropic was buying a chip-efficiency layer and the ~100 people who built it, not a product line. Decart was registered on September 7, 2023 by Dean Leitersdorf and Moshe Shalev, has raised about $450 million, and was valued at roughly $4 billion in May, per Calcalist. At $6 billion that was a ~50% step-up on that round, and about $60 million per employee.

The asset is DOS — the Decart Optimization Stack — which the company says "unlocks the full performance of every major chip" across NVIDIA, AWS Trainium and Google TPU hardware, according to Decart's own site. At the May round, Decart announced that DOS 2.0 runs agents at over 1,600 tokens per second — which its announcement put at eight times the industry average — and said it was generating significant revenue licensing DOS to cloud providers and AI labs. Treat both as vendor figures: they come from the company's own funding announcement, not an independent benchmark, and no third party has published a reproduction. Decart also ships consumer-facing world models — Lucy for real-time video transformation, Oasis for simulated environments — but Calcalist's reporting on the rationale puts the motive in cost terms: Anthropic, like its competitors, faces what it calls "the fundamental economic problem of frontier AI: enormous demand for computing comes with enormous costs," and extracting more work from the same silicon attacks that directly.

The most load-bearing detail was where the people would have landed. Bloomberg's August 13 report — the one carrying the $6 billion figure that turned out to be the durable one — said the Decart team would join Anthropic's inference and performance organization. Not research. Not product. The org whose entire job is making a token cost less to produce.

That team is now joining nothing. Anthropic gave no reason for walking and both companies declined to comment, so treat the cause as unknown rather than inferring one. What the process still tells you is where the company was aiming: nobody runs diligence on a $6 billion acquisition for a team they intend to file under research.

The Pressure That Produced the Bid Is Still There

Offering pre-IPO shares — Calcalist's characterisation of the consideration — said Anthropic rated the cost reduction above the paper it would have spent. Anthropic's annualized revenue run rate reached $65 billion at the end of July, up from $47 billion in May, against a $965 billion post-money valuation from its Series H — and the company has already filed confidentially for an IPO, with an internal projection of $190–200 billion in annual revenue by 2028.

Read those numbers together and the bid was coherent — and not one of them changed on September 8. A company about to be marked by a public market has one number it must move: gross margin on inference. It has locked in enormous fixed capacity — Anthropic, Google and Broadcom announced multiple additional gigawatts of next-generation TPU capacity on April 6, 2026, coming online from 2027, on top of existing Trainium and TPU commitments. Fixed capacity plus a software layer that gets more tokens per chip is the fastest available path to margin. Buying it with stock, before the stock is publicly priced, would have been the cheapest way to pay for it. None of that logic died with the deal; it is still on Anthropic's balance sheet, and it will look for another answer.

Anthropic was not the only bidder. Calcalist reported that NVIDIA — already a Decart investor from the $300 million May round — put a higher valuation on the table and the process moved to Anthropic anyway, with Google and SpaceX also named as potential buyers. Anthropic, NVIDIA, Google and SpaceX have all been named around the same 100 people. That is the market telling you what inference efficiency is worth right now. Decart ends the process holding neither the offer it took nor the higher one it passed on.

Anthropic's Price History Says Efficiency Reaches You Late

Cost reductions have historically reached Anthropic's customers as a new model number, not as a lower price on the model they already committed to. The evidence is on Anthropic's own pages.

Claude 3 Sonnet launched on March 4, 2024 at $3 per million input tokens and $15 per million output. Sonnet 4, Sonnet 4.5 and Claude Sonnet 4.6 are all still $3/$15 today, per Anthropic's pricing documentation. That is nearly two and a half years of flat list pricing across four model generations, through the largest efficiency gains in the industry's history. The streak broke only with Sonnet 5 in June 2026 — and it broke the way this whole section describes: on a new model number.

The Opus tier tells the other half. Claude 3 Opus was $15/$75 in March 2024; Claude Opus 4.5, announced November 24, 2025, came in at $5/$25 — a threefold cut. Anthropic framed the gain in tokens, not dollars: at its highest effort level, Opus 4.5 "exceeds Sonnet 4.5 performance by 4.3 percentage points—while using 48% fewer tokens."

Steel-man the vendor here, because there is a real and recent case. Claude Sonnet 5 launched on June 30, 2026 at $2/$10 — a 33% cut on the per-token rate Sonnet had held since 2024 — and Anthropic's docs record that the increase to $3/$15 scheduled for September 1, 2026 will not occur: the introductory rate became the standard rate. Prompt caching reads cost 10% of base input, the Batch API is a flat 50% off, and the full 1M-token context window carries no premium. Anthropic does pass savings through. It just does it on its own schedule, at the moment it ships a new model, to whoever is paying list — not to the customer who fixed a rate in a two-year agreement.

The Tokenizer Moved and Your Rate Card Didn't

A price per token is not a price, because the vendor defines the token. This is the part most procurement teams have not priced, and it is sitting in public documentation.

Anthropic's pricing page carries a note that Claude 4.7 and later models "use a newer tokenizer" that "produces approximately 30% more tokens for the same text," with the exact increase depending on content and workload shape. Sonnet 4.6 and earlier use the previous tokenizer. Now do the arithmetic on the vendor's own published figures: moving from Sonnet 4.6 at $3/$15 to Sonnet 5 at $2/$10 is a 33% cut per token, but the same body of text now bills roughly 30% more tokens. Multiply the two and the headline 33% price cut lands as roughly 13% on the same body of text. Text is not the same thing as work — a stronger model may finish a task in fewer passes — which is precisely why the corpus below, not the token, is the unit worth contracting on.

That is not an accusation of bad faith — a better tokenizer is a legitimate engineering choice, and Anthropic documented it. It is a demonstration that the unit is not stable, and a multi-year commitment denominated in dollars-per-million-tokens is a commitment to a unit your counterparty controls the definition of. The same trap runs the other way through effort settings, reasoning budgets, and per-model tool-use overhead: the pricing docs list a tool-use system prompt costing 286 tokens on Opus 5 and 675 on Opus 4.7 for the identical request. We have made this argument before about comparing frontier models on per-token price; the Decart deal is the reason it now belongs in the contract rather than the evaluation memo.

Two more multipliers stack on top and belong in any modelled rate: US-only inference via inference_geo carries a 1.1x multiplier on every token category, and regional or multi-region endpoints on Amazon Bedrock and Google Vertex AI add a 10% premium over global. If your data-residency posture is non-negotiable, your effective rate is 10% above the number on the page before you negotiate anything — the same structural surprise buyers hit with Mistral's regional European endpoints.

The Clauses That Survive a Price Collapse

Ask for a downward price-adjustment mechanism and a stable unit of work, in that order. Everything else is secondary.

Start from what the published paper actually says, because it is the floor you negotiate up from. Anthropic's Commercial Terms of Service provide that "Anthropic may update the published rates, to be effective the earlier of 30 days after the updates are posted by Anthropic or Customer otherwise receives Notice" (§H.1), and claude.com/pricing states plainly that price and plans are "subject to change at Anthropic's discretion." The same terms let Anthropic assign the agreement "to an affiliate or as part of a sale of all or substantially all its business" without your consent (§M.4). A standard change-of-control clause was never the lever here — the company in play was Decart, not Anthropic, and now neither is.

The four asks that matter:

1. A ratchet, not a rate. Your negotiated discount applies to list, and if list falls, your price falls with it. Never the reverse. A fixed dollar rate across a 24- or 36-month term is a bet against a vendor that just ran diligence on a $6 billion acquisition to lower its own cost per token.
2. Define the unit as work, not tokens. Pin a benchmark corpus — a few hundred of your real prompts and expected outputs — into the agreement, and define a price change as a change in cost to process that corpus. A tokenizer swap, a default effort-level change, or a mandatory reasoning budget that moves your corpus cost by more than a stated threshold is a price change and triggers the same 30-day notice and the same adjustment right.
3. Model-substitution at successor pricing. You commit spend, not a SKU. When the successor model ships, committed dollars move to it at its list price, without renegotiation and without your discount resetting. Otherwise the next Sonnet-5-shaped cut lands on everyone except you.
4. Drawdown, not prepay-and-forfeit. Unused committed spend rolls into the next period or converts. If efficiency gains mean you consume half the tokens you forecast, a use-it-or-lose-it commit converts your vendor's cost win into your write-off — exactly the repricing exposure buyers hit when Bending Spoons acquired Airtable and inference credits were re-cut mid-term.

Anthropic's own pricing page confirms this is negotiable territory: the sales-assisted Enterprise plan explicitly supports "MSA, PO, usage commitments, product bundling," and volume discounts are described in the pricing docs as negotiated case by case. Self-serve Enterprise is $20 per seat plus usage at API rates. If you are on the self-serve tier, you have no adjustment mechanism at all — you have a published price list and 30 days' notice.

If You License DOS, You Have a Different Problem

Decart's existing customers had a change-of-control question. They now have the opposite one, and it is not obviously the better one. SiliconANGLE's reporting says DOS revenue comes from licensing to cloud providers and AI labs. If that is you, the event you were bracing for did not fire: DOS is not about to be owned by a frontier lab you compete with, and the "hardware-agnostic" property you bought — spanning NVIDIA, Trainium and TPU — stays a product decision rather than one made inside a competitor.

What you have instead is a standalone counterparty that just burned its exit. Decart was valued at roughly $4 billion in May, passed over a higher NVIDIA number to run a process with Anthropic, and came out of diligence with no buyer — a public data point that sits under its next raise. Independence is not stability. It is a different failure mode, and it does not change the ask.

We wrote the general version of this when d-Matrix acquired Wallaroo: get "any hardware" in writing, with named platforms and a support horizon, before the deal signs. There is now no deal to sign before, which means nothing forces the conversation and you have to open it yourself. Ask for the support horizon across NVIDIA, Trainium and TPU by name, and ask what survives a change of control by any of the four buyers that have circled this company.

The consumer side is the same story. Lucy and Oasis are sold pay-as-you-go, priced per second by model — $0.02/sec realtime and $0.04/sec of generated 720p video for Lucy 2.5, $0.01/sec for Lucy Restyle 2, with no subscriptions and no minimum spend. "No minimum spend" is a nice consumer term and a terrible enterprise one: it means nothing constrains repricing — not by a future acquirer, and not by a company that has to fund itself alone for longer than it planned.

NVIDIA's position is still the detail to watch. It is simultaneously a Decart investor, the bidder that was passed over, and the vendor of most of the silicon DOS optimises. And the two companies that just spent a month in a data room may still find other ways to work together, per Bloomberg's sources — a commercial partnership between your efficiency vendor and a frontier lab raises a smaller version of the same question, and it will not arrive with a press release you can diary.

What to Do Before You Sign

This Week:

1. Pull every AI vendor agreement with a term longer than 12 months and grep for a price-adjustment clause. Most have an escalator and no ratchet. Write down which ones expire before Anthropic's expected listing window.
2. Freeze a benchmark corpus — 200 to 500 real production prompts with expected outputs — and run it against your current model. Record total input tokens, output tokens and dollar cost. That single number is your unit of account for every negotiation from here.
3. If a Claude commitment is mid-negotiation, do not argue from the deal — it is dead, and your account team already knows. Ask the question that outlived it: what happens to my rate when Anthropic's cost per token falls? Get the answer in the redline, not on the call.

This Month:

4. Re-run the benchmark corpus on the newest model tier and compare cost per completed task, not cost per token. If the token count moved more than the price did, you have found your clause.
5. Instrument per-model, per-task cost in your gateway so the corpus number is continuously measured rather than a one-off — the gateway comparison covers what to look for, and prompt caching hit rates matter more than routing here, as we found when caching broke the multi-model router economics.
6. Inventory whether anything in your stack depends on Decart — directly, or through a cloud provider or model vendor that licenses DOS. The acquisition failing does not close this item; it leaves the dependency sitting under a company that just lost its exit. Vendor-of-your-vendor exposure is the disclosure gap that keeps recurring, most recently with Stripe's acquisition of OpenRouter.

Before Renewal:

7. Refuse any fixed per-token rate longer than 12 months without a downward adjustment mechanism. Take a shorter term at a worse headline discount over a long term at a frozen rate. In a market where your vendor was willing to spend $6 billion of its own equity on cost reduction, optionality is worth more than three points of discount.

The Bottom Line

This is the reserved-instance trade, replayed with a worse-defined unit. Cloud buyers learned in the 2010s that a three-year commitment at today's rate is a bet the provider's costs will stop falling — and the providers kept shipping cheaper instance families to everyone except the people who had prepaid. The AI version is harder, because at least a vCPU-hour meant the same thing in year three as it did in year one. A token does not: Anthropic's own documentation says its newer tokenizer produces about 30% more of them for the same text.

The company that committed multiple gigawatts of TPU capacity from 2027, and was willing to put roughly $6 billion of pre-IPO equity behind a chip-efficiency team, told you with its capital that it expects to produce a token far more cheaply than it does today. On September 8 it walked away from that particular team. It did not walk away from the expectation. Believe the expectation — then make sure your contract is written in a unit that lets you collect on it.

Price the work. Never the token.

Continue Reading

• Claude vs GPT vs Gemini: Stop Comparing Per-Token Prices
• d-Matrix Bought Wallaroo. Get 'Any Hardware' in Writing.
• Bending Spoons Bought Airtable. Reprice Before It Closes.
• Mistral Wants Multi-Year Money. Its EU Endpoint Drops Agents.
• Microsoft's GPUs Sit Unplugged. Buy Delivery, Not Capex.
• DeepSeek Will Raise Prices. Your Ceiling Is Already 4x.
• AMD Bought Taalas. Now Name the Model You'd Freeze.
• Stripe Bought OpenRouter. A Toggle Is Not a Contract.