Nvidia wants to buy the "American DeepSeek." A Delaware courtroom just showed us what that kind of deal costs the people who built the technology. Both stories broke within a week of each other, and together they tell you everything about how the AI land grab actually works now — not through press releases, but through licensing structures, equity side-doors, and lawsuits alleging that billions went where the paperwork said they shouldn't.
On Saturday, October 10, the Financial Times reported that Nvidia is in early talks to either acquire Reflection AI outright or dramatically deepen its existing stake in the startup — the open-weight AI lab that Washington-friendly voices have spent a year branding as America's answer to DeepSeek. Reflection was last valued at $25 billion. Nvidia already owns $800 million of it. Per the FT, a deal could land "within weeks," though talks at this stage can absolutely fall apart. Reuters, for the record, could not independently verify the report — keep that asterisk handy.
Reflection AI was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both veterans of Google DeepMind, with a pitch that was geopolitical from day one: build frontier-grade open-weight language models so the open-source tier of AI doesn't get ceded to Chinese labs like DeepSeek. That positioning got a whole lot louder in January 2025, when DeepSeek's R1 release rattled Washington by proving an open-weight Chinese model could rival American closed models at a fraction of the training cost.
The money followed the mission. In October 2025, Nvidia wrote an $800 million check as part of a $2 billion raise that valued Reflection at $8 billion. By the April 2026 round, that valuation had hit $25 billion pre-money — Laskin confirmed the number to CNBC himself. Triple your valuation in under a year, and people start asking whether renting influence is still cheaper than owning it.
Then came the product. On October 5, Reflection launched Beam, its first open-weight model: a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens with a 1-million-token context window. The company claims Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while using three to four times less inference compute — and, in a detail that tells you who Beam's real audience is, it runs on Nvidia silicon. Full weights are promised under an Apache 2.0 license later this month. Two caveats before you take the benchmark slide to your board: the efficiency claims haven't been independently verified, and by Reflection's own published benchmarks, Beam still trails DeepSeek and Kimi K3 on several measures. "America's answer" is currently grading its own homework.
Per the FT and Bloomberg reporting, three options are on the table:
The FTC is not asleep. Chair Andrew Ferguson said back in January that the agency is "beginning to examine these acqui-hires to make sure they are not an attempt to get around" the merger review process. U.S. senators put a name on the pattern in February, labeling deals like Nvidia-Groq "reverse acqui-hires." The FTC has been here before — its scrutiny of Microsoft's handling of Inflection AI's team in 2024 turned this specific structure into the most-watched maneuver in tech M&A. An acqui-hire can avoid some filing requirements. It does not eliminate antitrust scrutiny, and it clearly has not eliminated litigation risk either.
Here's the part of the Reflection story that should stop you cold: training Beam involved a four-week reinforcement-learning run across 10,500 Nvidia GB300 GPUs, rented through a $6.3 billion, multi-year compute deal with SpaceX's Colossus 2 facility — Reflection paying roughly $150 million a month through 2029. Weeks later, Reflection signed a separate $1 billion-plus compute deal with Nebius, again for Nvidia GPU access through 2029.
Do the arithmetic. Reflection burns nine figures a month, and every one of those dollars lands back on Nvidia's income statement. Nvidia collects rent whether Beam wins or loses. If Reflection stumbles, Nvidia keeps the compute revenue and writes off an equity position it can afford to lose. If Reflection wins, Nvidia holds an $800 million ticket on a $25 billion company — and a strategically priceless asset in the open-weight lane.
Can Nvidia afford the full buyout? Trivially. The company generated about $127 billion in free cash flow over the past four quarters — a $25 billion check is roughly ten weeks of cash generation — and held $56.6 billion in cash and equivalents as of July 26. And this isn't an isolated pattern: Nvidia is separately in talks to invest in Perplexity at a valuation above $30 billion, up more than 50% in a year. The blunt version, per Startup Fortune's read: Nvidia is done being just the arms dealer and wants a cut of the war itself.
Now the story Steve flagged — and where the popular framing needs correction, because the viral summary and the actual court filing are two different things.
What's true: On October 2, two former Groq engineers — Joshua Rubin and Benjamin Serebrin — filed suit in Delaware's Court of Chancery (unsealed October 5). Both left Groq before the December 2025 deal was announced, but held stock. The proposed class action alleges Nvidia's $20 billion transaction with Groq "squeezed out" stockholders at a "lowball" price.
Correction one — who's being sued: The defendants are Groq's former board and former CEO Jonathan Ross. Nvidia is not a named defendant in this action. The lawsuit is a fiduciary-duty case about how Groq's own board ran the sale, not a damages claim against Nvidia.
Correction two — what's being alleged: The complaint does not argue the license fee was too small. It argues the opposite problem — that the money was structured to route value past most shareholders. Per CNBC's review of the filing: Nvidia allocated $17 billion to a license labeled "non-exclusive" and set aside $3 billion in Nvidia restricted stock units for the roughly 150–200 Groq engineers who transferred along with the technology. The suit alleges the board approved the deal without the stockholder vote Delaware law requires, ran no process "designed to test or maximize the value of what Nvidia bought," and that a board majority was conflicted because investment funds that designated Groq board members were "positioned to enjoy windfall returns from the later squeeze-out." The plaintiffs' own words: "The Board's conflicted choice cost Groq's stockholders billions of dollars."
The accounting receipt: Nvidia's own 10-K is the most damning exhibit in the plaintiffs' favor. Nvidia booked $14.4 billion of the payment as goodwill — "primarily attributable to the workforce and future development of the licensed technology" — and just $2.5 billion for the technology itself, with notes stating "no customer contracts, existing products, or equity interests were purchased." In plain English: roughly 85% of what Nvidia paid was, by its own accounting, a bill for people — not a license for code. That's precisely the fact pattern that makes a "licensing deal" look like an acquisition wearing a licensing costume, and precisely why the plaintiffs concede no Delaware court has directly answered the question this case raises.
The counterweights: Groq called the lawsuit "meritless" and said the agreement "delivered exceptional value for Groq, our investors, and our employees." Axios reported in December, citing sources close to the deal, that most shareholders would receive per-share payouts tied to the $20 billion valuation, with vested shares paid in cash and unvested shares converted to vesting Nvidia stock. There's also a reported DOJ inquiry into the deal — a formal request for information sent in September, per the New York Times, with a fine possible — which cuts the other direction: regulators are asking the same structural questions the lawsuit does. Nvidia's response at the time was, charitably, confident: the Groq deal was "a prime example of the American system working as designed."
One more twist that shows how strange this deal is: Groq still exists. It raised a $350 million Series A in August at a $3.5 billion valuation (down from $6.9 billion pre-deal), pivoted entirely to AI cloud computing, and is now deploying "NVIDIA Groq 3 LPX" racks — built on the very LPU technology it licensed to Nvidia. The technology's former owner is now its first customer class. Whether that's elegant restructuring or hollowing-out depends on which side of the complaint you're reading.
I'll give you the BDM read, because that's my job.
Nvidia's neutrality is ending. The moment the company that sells you GPUs also owns or employs the model layer, every AI lab with a pulse — OpenAI, Anthropic, Google, Mistral — has to ask whether its supplier is becoming its competitor. Nvidia already funds OpenAI. It's circling Perplexity. It employs Groq's LPU team. It may own the "American DeepSeek." If you're a startup building on Nvidia silicon and competing in the model layer, you now have a supply-chain concentration risk your investors will ask about in diligence.
The acqui-hire is now a contested structure, not a clever one. Groq was the template. The template now has a class action, a DOJ information request, FTC chair comments, and Senate attention attached to it. Expect Reflection deal-structure language to be lawyered to death — and expect more "full acquisition" outcomes like Hugging Face, because clean deals don't generate subpoenas.
Watch the structure, not the price. $25 billion is a rounding error for Nvidia. Whether Reflection arrives as a purchase, a license-plus-hire, or a bigger stake tells you how much regulatory heat Nvidia thinks it can carry — and that signal matters more than the number itself.
Open-weight geopolitics is now an M&A thesis. Huang said in July that the US must build its own open AI ecosystem. Buying the leading American open-weight lab is the fastest possible way to make that true — and to make sure the open-weight future, as Startup Fortune put it, "doesn't have to run through Beijing."
The 快科技 headline Steve sent — "NVIDIA negotiating to buy the American DeepSeek" — is directionally right. The details are where it gets interesting. Nvidia is negotiating three different deals at once, and the one it picks will tell the market whether the acqui-hire era is over. Meanwhile, the Groq lawsuit is the fine print arriving early: when a $20 billion "license" books 85% as goodwill for people, workers will eventually ask whose value that goodwill represented — and Delaware's chancellors will eventually answer.
Structure is the story. Watch which door Nvidia walks through.
