Nvidia paid $6 billion for the machine Poolside used to build its models, hired the 109 engineers who ran it, and left the company legally untouched. No control changed hands. No entity was absorbed. The three co-founders are still in their chairs, and Poolside is still selling you the same product under the same paper.
That is the problem. If you run Poolside's models inside your own perimeter — and its buyers are banks, defense primes, intelligence programs and federal offices that cannot let code leave the building — then every continuity protection you negotiated is keyed to ownership, and ownership did not move. Your change-of-control clause did not fire. Your anti-assignment clause did not fire. Your M&A notification clause did not fire. Meanwhile the team that built the model, the system that builds models, and the vendor's stated reason to keep building them all left on the same Thursday.
What Nvidia Bought, and What Stayed Behind
Nvidia bought the capability, not the company — and the distinction is deliberate.
According to a letter to Poolside's investors first obtained by Newcomer and reported this week, Nvidia is paying $6 billion for a non-exclusive licence to the "Model Factory" — the system Poolside used to build its Laguna family of coding models — and is making job offers to 109 Poolside employees who worked on Laguna. Separately, Nvidia is investing $1 billion at a $12 billion pre-money valuation, which Dealroom puts at $13 billion post-money, with the $6 billion licence fee expected to reach Poolside's investors by the end of 2027.
Poolside's letter is explicit about what this is not. The arrangement is, in the company's own words, "not an acquisition and it is not an acquihire". That is legally accurate. It is also the whole point.
Consider the scale of what did move. Third-party headcount estimates for Poolside vary too widely to build an argument on — from about 150 to 264 depending on the tracker — so use the co-founder's own numbers instead. Describing the team behind the model, Eiso Kant said "Less than 70 people built this model... Less than 115 between engineering and researchers, like, together did this effort". Against that denominator, 109 offers is not a slice of the company. It is substantially the entire organisation that produced Laguna S 2.1, a 118-billion-parameter mixture-of-experts model with 8 billion active parameters, built on 4,096 Nvidia H200 GPUs in under nine weeks and scoring — on Poolside's own published benchmark table — 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual, with weights published on Hugging Face under a permissive licence.
The licence is non-exclusive, which sounds like protection and is not. Poolside retains the right to license the Model Factory to Google, Amazon or anyone else. What it does not obviously retain is the 109 people who knew how to operate it.
Why Every Continuity Clause Stayed Silent
Your contract stayed silent because standard continuity language is keyed to ownership, and this transaction moved everything except ownership.
A change-of-control clause is a provision that triggers rights — consent, notice, termination — when a counterparty is acquired or its ownership changes. In practice the definition turns on one of three events: a threshold of voting power, a merger or consolidation, or a sale of all or substantially all assets. Run the Poolside deal against all three:
- Voting power. A $1 billion investment at $13 billion post-money is a single-digit minority stake. Nowhere near a control threshold — and financing rounds are routinely carved out of the definition by negotiation anyway.
- Merger or consolidation. There is none. The company continues to exist as the same legal entity, with the same founders.
- Sale of substantially all assets. A non-exclusive licence is not a sale. Poolside still owns the Model Factory and can license it again tomorrow.
Job offers to 109 engineers appear in none of those tests, because employment is not an asset a company can convey. That is why a deal of this size can close without a consent right anywhere in your agreement waking up.
The same logic explains why the licence itself required no premerger filing, which is the part that should worry a procurement lead most. The 2026 Hart-Scott-Rodino premerger notification threshold is $133.9 million in transaction value, effective 17 February 2026, and it applies to acquisitions of voting securities, assets, or non-corporate interests. A $6 billion acquisition clears that bar forty-five times over. A $6 billion licence does not clear it at all, and the reason is exclusivity. The agencies treat the grant of an exclusive licence as the transfer of an asset — but for that treatment to apply, "the license must be exclusive – even against the grantor". Poolside's licence is expressly non-exclusive and Poolside can grant more tomorrow, so no asset moves. Add a hiring campaign, which is not an acquisition of anything, and the structure, as one analysis put it, "sidesteps acquisition scrutiny" by design. Nvidia's separate $1 billion equity purchase is its own transaction on its own facts; the licence and the hiring — the two things that actually moved the capability — are what pass unexamined.
Steel-manning the other side: nothing here is unlawful, and there is a real argument that the customer is better off. Poolside gets a balance sheet it could not otherwise raise, its investors get liquidity, and the models you already run keep running under a permissive open-weight licence you can host yourself. That argument has force. It also has nothing to do with whether your contract gave you a decision point, and it did not.
Nvidia Has Run This Play Three Times in a Year
Treat this as a repeatable structure, not a one-off, because it now has a track record.
In September 2025, Nvidia paid roughly $900 million to license Enfabrica's networking technology and hire its CEO, with the company left standing. In December 2025 it went far bigger: a reported $20 billion non-exclusive licence to Groq's LPU architecture, compiler toolchain and silicon design work, with founder Jonathan Ross and president Sunny Madra moving to Nvidia while Groq continued as an independent company under a new CEO and kept GroqCloud running — the structure confirmed in Groq's own announcement, even as headlines described Nvidia as "buying Groq's assets." Poolside is the third instance in eleven months.
None of this is unique to Nvidia. The pattern was already established in 2024, when the UK's Competition and Markets Authority opened a probe into Microsoft's $650 million licensing deal with Inflection and the FTC asked questions about Amazon's arrangement with Adept — precisely to test whether the structure was sidestepping rules a straight purchase would trigger. The CMA cleared that arrangement in September 2024. Two years on, no regulator has stopped the structure — Nvidia has used it three times since.
That is not the same as regulatory indifference, and the counter-argument deserves stating. FTC chair Andrew Ferguson said in January 2026 that acqui-hiring "has become a big enough deal" that the agency is "beginning to look very closely at how these things work"; the DOJ Antitrust Division's acting head has called the structure a red flag; and in February 2026 a group of senators asked both agencies to treat these deals as "de facto mergers" that "bypass the scrutiny typically applied to mergers and acquisitions." Former FTC commissioner Josh Wright goes further in the other direction, arguing the whole evasion narrative is "dramatically overstated" — agencies review unfiled transactions routinely, he reports they examined the Nvidia-Groq deal, and the Clayton Act lets them challenge a transaction after it closes. He may well be right. It changes nothing for you. An agency review is not a notice to customers, it runs on a timescale of years, and it produces no decision point in your contract. Even on the most optimistic reading of the regulatory posture, the buyer finds out from the press.
The practical conclusion for a buyer is uncomfortable and simple. The most likely form of your next AI vendor continuity event is not an acquisition. It is a licence, a hiring wave, and a press release saying nothing has changed. We have covered the ownership-transfer version repeatedly — Tricentis buying Tabnine's context engine, Stripe buying OpenRouter, Klaviyo winding down Agency's product. In every one of those, a clause fired and somebody got a letter. Here, nobody did.
Read Poolside's Letter as a Capacity Disclosure
The letter says out loud why the deal happened, and the reason is the thing your contract should have been tracking all along: compute.
Poolside told investors it "had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn't close it in time, and we lost the cluster" — and that continuing to compete in model development would have required more Nvidia hardware than it could secure.
Set that against where the company was ten months earlier. In October 2025, CoreWeave announced it would be anchor tenant for Project Horizon, a 2GW campus in West Texas with a first phase of 250MW and more than 40,000 GB300 NVL72 GPUs. Co-founder and co-CEO Eiso Kant said in that announcement: "To compete at the frontier you need to be vertically integrated from dirt to intelligence." Ten months later, the frontier ambition is a licence fee payable to the company that makes the chips.
This matters to you specifically because of what Poolside sells. Its public-sector offering advertises air-gapped and on-premises deployment with full model weights, ATO achieved, STIG-hardened OS support and IL5 deployability, with named partners including Northrop Grumman, Dell, Cubic, Atos and IQT. Its enterprise path runs single-node RKE2 on Dell hardware, Helm on customer Kubernetes, or bare metal on your own Nvidia infrastructure, down to edge boxes like the NVIDIA DGX Spark. Those are exactly the deployments that cannot fall back to somebody's API next quarter. When a vendor sells you sovereignty, its ability to keep producing models is the product — and that ability is now a question, not an assumption. It is the same trap as the Army's IL5 Agentforce authorisation shipping with the models switched off: the platform is fine, the intelligence inside it is the variable.
The Three Triggers Your Contract Is Missing
Write continuity triggers against capability, not against the cap table. Three clauses do almost all the work.
1. Key personnel and capability continuity. Federal contracting has had this language for decades. Under the HHS acquisition regulation, a contractor must notify the contracting officer at least 30 days before voluntarily diverting named key personnel, and "shall not divert, replace, or announce any such change to key personnel without the written consent of the Contracting Officer". Commercial AI contracts almost never carry an equivalent. Name roles, not individuals — "the engineering leadership responsible for pretraining and post-training of the licensed model family" — with a notice obligation on departure above a stated threshold and a termination right if replacements are not proposed.
2. Core-IP licensing. A notification and termination right that fires when the vendor grants any third party a licence — exclusive or non-exclusive — to the model architecture, training pipeline or model-production system underlying your deployment. This is the clause that would have fired last Thursday. It fires on the movement of the thing that makes the product, which is the only definition that survives a structure designed to leave ownership alone.
3. Material change in the vendor's business. The event you actually care about is "we are no longer developing frontier models," and no ownership test detects it. Define it: an announced exit from a product line, a stated inability to secure the compute the roadmap requires, or the loss of a named production dependency.
Behind all three sits escrow, and AI escrow is not source-code escrow with a new label. Standard release triggers are vendor insolvency, material breach of maintenance obligations, and product discontinuation — and for an AI deployment the deposit has to include trained model weights, training data pipelines, hyperparameter configurations and inference runtime dependencies, or you have archived, in that guide's phrase, "a car engine without the fuel system." Then check that yours mandates verification — that somebody has confirmed the deposit is complete and can actually be rebuilt. It is the clause most often left out, because it is the only one that costs money before anything goes wrong, and an unverified deposit is a promise rather than a control. Poolside's open-weight releases blunt this particular case; your next vendor's closed weights will not.
What to Put in the Next Renewal
This Week:
- Separate your AI vendor list into two columns: vendors whose models you run inside your perimeter, and vendors whose API you call. Column one is where a personnel-and-licensing event is existential; column two is where you can switch in a sprint.
- Pull the top five contracts from column one and read the definition of change of control, not the clause that references it. If the definition is voting power, merger and asset sale, you are unprotected against this structure. Search the same documents for "key personnel," "assignment," and "material adverse change."
- Count how many of those five have an escrow arrangement, and how many of those escrows specify model weights. Write the number down; it will be lower than you expect.
This Month:
- Send three questions in writing to any vendor in column one, Poolside included: which named engineering functions remain in place; whether any third party now holds a licence to the model-production stack; and what the roadmap for the next model release is, with a date.
- Draft the three triggers above as standing redlines owned by procurement, not re-argued per deal. One paragraph each.
- Run a capability-loss tabletop with your engineering lead: the vendor ships no new model for 18 months. What breaks, what do you pin, and what is the migration path? Use your existing platform scoring — the exit criteria approach we laid out for agent orchestration transfers directly.
Before Renewal:
- Escrow with mandatory verification and an explicit deposit manifest: weights, tokenizer, training and fine-tuning pipeline, hyperparameters, inference runtime.
- A continuity notice obligation with teeth — 30 days, written, with a defined termination window — modelled on the federal key-personnel language rather than invented from scratch.
- A named fallback in the contract itself. If you are self-hosting an open-weight coding model, the fallback is another open-weight model you have already evaluated; if you are on a commercial assistant, know what a move to GitHub Copilot or a rival actually costs in migration weeks. We scored that field at 500 seats here, and the sovereignty-constrained options here.
The Bottom Line
Enterprise software contracts learned to watch ownership because, for thirty years, ownership was where the risk lived: a rival bought your vendor, the roadmap died, the price list changed. AI has moved the risk somewhere the paperwork does not look. The scarce asset is not the company — it is a few hundred people who know how to train a frontier model and the compute contract that lets them do it. Both can leave without a single share moving, and this deal is the proof.
The right response is not outrage; there is nothing to be outraged about. Poolside got a balance sheet, its investors got liquidity, and Nvidia got the production system rather than the output. The response is to stop writing clauses that only fire when a company is sold.
Ownership is the easiest thing to watch and the least likely thing to move. Write your triggers against the people and the machine, not the cap table.
Continue Reading
- Tricentis Bought Tabnine's Context Engine, Not Your IDE
- Your AI Vendor Joined OpenAI in March. You Heard in August.
- Stripe Bought OpenRouter. A Toggle Is Not a Contract.
- Groq Runs Nvidia Now. Recount Your Non-Nvidia Capacity.
- d-Matrix Bought Wallaroo. Get 'Any Hardware' in Writing.
- Anthropic Bid $7B for Cheaper Inference. Don't Fix Your Rate.
- Oakley Bought Graphwise. GraphDB Isn't Open Source.
- Best AI Coding Assistant at 500 Seats: Buy Copilot Business