Nvidia Reflection AI Acquisition Talks|$25 Billion Price of Keeping Chip Buyers Many?

Why a chip seller wants a model maker

Nvidia is in talks to buy Reflection AI, an American open-model startup, or deepen its stake, the Financial Times reports. The company that sells chips to AI labs is weighing whether to own one, valued at 25 billion dollars. The paper said the talks are early and could still fall apart. Nvidia and Reflection did not immediately respond to requests for comment.

Nvidia has already invested 800 million dollars in Reflection, the FT reported. My read is this. Nvidia is paying to keep the buyers of its chips many, rather than concentrated in a few giant labs. That is a claim about Nvidia's pricing power, and the evidence will test it. Reflection's own pitch seems to cut against that. Its first model, Beam, was unveiled on October 5.

Reflection says Beam scores on par with China's GLM-5.2 on advanced reasoning tests. It says Beam does that with three to four times less inference compute. Those claims have not been independently verified. Still, the selling point is plain. It is a model built to need fewer chips, backed by the dominant supplier of AI processors. The answer starts with what open-weight means.

Developers can download the model's parameters and run them on their own hardware. Reflection's chief executive, Misha Laskin, told Semafor it sees demand from groups unwilling or unable to use Chinese models. Reflection is also a heavy renter of Nvidia hardware. It reportedly signed a 6.3 billion dollar deal with SpaceX to rent Nvidia GB300 systems, and used them to train Beam.

Reflection says training ran on 10,500 Nvidia GPUs over four weeks. It also signed a one billion dollar deal with Nebius for Nvidia's latest chips. The FT, citing Ramp data, reports businesses shifting to cheaper models that get the job done. Analysts like Gartner warn total inference bills may keep climbing anyway. So I'd read Beam's efficiency as widening who runs AI on their own compute.

For a chip seller, that means more buyers. Reflection is not Nvidia's only move toward open models. In September, Nvidia agreed to buy Hugging Face, the main platform for sharing open models. The price was 12.9 billion dollars. Reflection was last valued at 25 billion dollars, nearly twice that. Nvidia also agreed to invest in Nous Research, backing both open-weight and proprietary model work.

And Jensen Huang has called for an open-weight "ecosystem," according to reporting on the talks. Hugging Face co-founder Thomas Wolf described the logic. He said: "We think the future of AI hopefully will be more than just two companies selling tokens to everyone." An analysis on Yahoo Finance names those two as OpenAI and Anthropic. It argues that if they sell most AI, they also control most of the demand for chips.

Many model buyers, it says, leave Nvidia with more pricing power than two big sellers would. That pricing power shows in the numbers. Nvidia's gross margin was 75 percent in its fiscal second quarter of 2027. That margin is what a fragmented market of model makers protects. The same analysis adds a catch, citing Ramp data. Open models already produce a large share of AI tokens.

Yet they draw under 5 percent of what businesses spend on AI. So Nvidia is paying up in the slice that earns the least today. In my reading, that makes it a bet on future share, not current revenue.

Why own instead of fund

Nvidia already funds labs without owning them. OpenAI is a major Nvidia customer, and has received a 30 billion dollar investment from the chip giant, CNBC reported. Before Nvidia's Hugging Face deal, OpenAI tried to invest 100 million dollars in Hugging Face, sources told CNBC. Under that plan, Hugging Face would have distributed OpenAI's own custom chips. OpenAI is making those chips with Broadcom.

According to CNBC's sources, Huang has privately complained about OpenAI's move into chips. AMD was also in talks with Hugging Face. The deal landed with Nvidia. Huang told CNBC open-source AI "was at the turning point." That case reshapes my opening claim. A lab can take Nvidia's money, buy its chips, and still build chips of its own. OpenAI does all three.

In my reading, a stake alone does not decide whose chips a lab favors. The goal looks narrower. It is many buyers, running on Nvidia. An analysis on Yahoo Finance reads the Nous deal the same way. It says Nvidia's backing makes enterprises more likely to run on Nvidia hardware and software, whichever models they adopt. Apply that to Reflection.

It has raised close to 2.6 billion dollars, from backers including Nvidia, Sequoia Capital and Lightspeed. Nvidia's 800 million dollars is roughly three of every ten of those dollars. That is a major backer's seat, the kind Nvidia also holds at OpenAI. Reflection's compute arrives through SpaceX and Nebius, not from Nvidia directly. A deal would give Nvidia a firmer hand in where Reflection's next models get built.

Who pays for the strategy

The Wall Street Journal has flagged a quiet change in how Nvidia counts free cash flow. Free cash flow is usually operating cash flow minus capital spending. Investors treat it as cash a company can return to shareholders or use to pay down debt. The Journal argues strategic equity stakes are capital spending too, since that money is tied up driving growth.

It says Nvidia subtly endorsed that view, though few people noticed. The upshot, per the Journal: Nvidia still plans big buybacks and dividends, but only after its investment needs. In late September, Nvidia's board added a record 150 billion dollars to its buyback authorization. That brought total repurchase capacity to 235 billion dollars through fiscal 2028.

In its fiscal second quarter, Nvidia returned roughly 26 billion dollars through buybacks and dividends. Reflection's last valuation, 25 billion dollars, is close to one full quarter of those returns. Under the Journal's reading, stakes like Reflection get paid for before that cash reaches shareholders.

Critics call the pattern circularity: capital recycled among companies that are each other's investors, customers and suppliers. Financial Times reporter Robin Wigglesworth put it bluntly to CNBC. "AI startups are burning through investor cash and handing it straight to Nvidia to buy chips," he said. Reflection's deals trace a version of that circle. Nvidia put 800 million dollars into Reflection.

Reflection signed a one billion dollar compute deal with Nebius, which provides Nvidia chips. Nebius itself took a two billion dollar investment from Nvidia. Add the reported SpaceX deal, and Reflection's compute commitments come to about 7.3 billion dollars. That is roughly nine times Nvidia's stake. That money flows to SpaceX and Nebius first, not straight to Nvidia. This is where my claim is shaken.

Part of the demand Nvidia protects is demand it helps finance. CNBC notes the AI startups and labs buying Nvidia's compute still can't prove standalone revenue. Analyst Dan Ives sees it differently. He argues Nvidia's results put several major worries to rest, with circular financing among the concerns. But chip sales show what buyers spent, not whether they can pay their own way.

So the claim survives in a narrower form. Nvidia is buying time for open-model builders to earn their own money. And shareholders pay for that time first.

What a deal would actually buy

The FT headline says acquire. The reported options are wider. A deal could be a so-called acqui-hire, where Nvidia hires staff and licenses technology instead of buying the company. That could avoid a lengthy regulatory review, the FT said. It could also be more equity, or a deal to supply more chips and computing power. Deal terms were not available. Benzinga notes Nvidia took the acqui-hire route with Groq.

In December, AI chip startup Groq announced a licensing deal with Nvidia, reported as a 20 billion dollar deal for its assets. Groq's founder and its president joined Nvidia. Groq said it would carry on as an independent company. That structure is now in court. Two former Groq engineers who held stock sued in Delaware.

They allege Groq's board "sold the company to Nvidia without the stockholder vote Delaware law requires." Groq's spokesperson said the licensing agreement "delivered exceptional value for Groq, our investors, and our employees." The spokesperson called the suit "meritless." Neither statement settles the price. What the case shows is the structure's weak point.

Investors left inside the startup can argue they were bypassed. Reflection has outside investors who would sit in that position. Besides Nvidia, they include Sequoia Capital, DST Global, Lightspeed and 1789 Capital. An acqui-hire is not invisible to regulators either. Apple disclosed its acqui-hire of audio startup Huxe in a European Commission filing.

That deal gave Apple hiring rights and a non-exclusive license, not the company. So the form decides what Nvidia shareholders end up owning. An acqui-hire buys people and a license. A full purchase starts from a 25 billion dollar valuation, Benzinga notes. Either way, the deal's shape will show whether Nvidia wants to fund open models or own one.

Back to the opening tension: a chip seller weighing whether to own a model maker. My read is now narrower than where I started. Nvidia appears to be paying to keep open-model building numerous, and running on its hardware. The Hugging Face contest and OpenAI's chip plans help explain why ownership, not just funding, is on the table. Shareholders fund that first, from cash that could otherwise go to buybacks.

On what's reported so far, the weight leans toward defending pricing power, with a real bill attached. The signal to watch is Beam's open weights. Reflection has promised downloadable weights under an Apache 2.0 license later in October. If they ship, and stay open under any Nvidia deal, this read holds. If they are held back or restricted after a deal, the funder part of this read breaks.

Nvidia would own a model product, set against closed labs like OpenAI, a major Nvidia customer. Organizations that wanted a Western model to run on their own hardware would lose that option.

Informational only, not investment advice. Figures and quotes come from the linked reports.