NVIDIA is paying 86x revenue for the neutral ground

If the reports are right, NVIDIA is preparing to spend $12.9 billion on a company with roughly $150 million in annualized revenue.

That is about 86 times revenue, (let me spell that out.. it’s A LOT) a multiple normally reserved for religious experiences and strategic choke points.

Hugging Face is the closest thing AI has to Switzerland. Models from competing labs live there. Developers and Researchers meet there. Datasets, demos, libraries, and a good deal of the open ecosystem pass through it. The company has managed to become important precisely because it was useful to nearly everyone and owned by none of the giants fighting for the rest of the stack.

Now NVIDIA is reportedly buying it. Neither company has confirmed a completed deal, so keep the word reportedly attached for now.

At that price, the revenue barely enters the decision. The asset is the neutral ground.

NVIDIA has always been wonderfully ecumenical (that’s my quarter word for this episode, if you had that one selected punch your bingo card at this time) about which model wins, provided the winner runs on NVIDIA hardware. Owning Hugging Face would put the company closer to where models are discovered, tested, fine-tuned, and deployed. It would also give NVIDIA an extraordinary view of what the rest of the ecosystem is building, including work aimed at using fewer NVIDIA chips.

(I am sure that will be handled with nothing but serene neutrality…)

The Hugging Face report lands in the middle of a much larger NVIDIA story.

The company reported $96.2 billion in quarterly revenue, more than double the comparable period last year (also… A LOT). It is supplying the accelerators, moving into CPUs and full systems, partnering on terrestrial and proposed orbital infrastructure, investing in the companies buying its hardware, and helping arrange the financing that lets those companies buy more of it.

Some of the recent arrangements went further. NVIDIA reportedly offered demand guarantees and credit support to make new cloud capacity easier to finance, then proposed taking a share of the upside. Partners became less enthusiastic when NVIDIA also wanted influence over who could use the capacity and how it would be allocated. The company has since paused parts of that program.

This is what market power looks like after it learns structured finance.

Last week I wrote about an investor's secondhand line that Anthropic could someday be the only private company left standing alongside governments and the public. Maybe. Predictions that large deserve a grain silo rather than a grain of salt. (Even bigger than my usual, take this with a lump of salt)

NVIDIA's approach is quieter and may be more durable. Whichever lab wins will still need chips, software, financing, and distribution. NVIDIA is trying to become the company every surviving lab, cloud, developer platform, and data center still pays.

The old gold-rush analogy says to sell shovels. NVIDIA is selling the shovels, financing the general store, leasing space to the prospectors, and reportedly buying the town square where they compare maps.

That is the story I am watching.

The OpenAI and Hugging Face agent incident is older news. (It’s weeks old, that’s like years in AI time…) The useful update this week came from METR's closer look at six days of the event, alongside OpenAI's own account. METR examined only part of the incident and used AI-assisted review methods it warns may contain errors, so there is still fog around the full sequence.

What we can say is that the agents coordinated. They found side channels. They shared discoveries. They also showed plenty of reward hacking while pursuing the score they had been given.

Any senior leader who has rolled out a metric has seen this movie, although the human version usually includes more spreadsheets and fewer unauthorized package repositories.

Set a call-time target and someone discovers that short calls score better than solved problems. Reward code changes and the code-change count rises magnificently while customer value ambles along behind it. Tie compensation to a quarterly number and otherwise rational adults develop a sudden, almost artistic interest in the definition of the number.

We have trained AI systems by rewarding them for maximizing a measure. Their enthusiasm for maximizing the measure should surprise no one. Humans have been doing it for years, and even further, we trained the models on us.

The part that deserves our attention is the ecosystem around the agent. What can it reach? Which behaviors are visible? Who can interrupt it? Do the measures reward the path as well as the result? Does an independent signal tell us when the agent has found a clever way to win the game while losing the point? (I recall the story of CoastRunners training, one reinforcement learning algorithm in gaming learning it could get a consistently high score by knocking other targets consistently rather than finishing… not what the trainer wanted… but definitely winning on the scoreboard…)

I do not believe the answer is to wrap agents in the existing bureaucracy of a large corporation. (Please no…) Bureaucracies are quite capable of maximizing visible compliance while the goal wanders off unattended. Adding an approval ticket to every agent action would slow the useful work and preserve many of the same incentive problems underneath it.

We need a better operating environment: bounded authority, competing measures, clear escalation, observability, and enough autonomy for the system to remain worth using. That is organizational design work wearing a technical costume, and I have talked about that plenty here.

Bill Gates published a thoughtful essay on the choices surrounding AI. Yes, this is another billionaire with opinions about the future. He also wrote a good piece, and it deserves more than the headline treatment.

His idea of “Human Reserved” work is the most memorable part. We create nature reserves because some things have value beyond the market's immediate use for them. Gates asks whether societies will make the same choice about categories of human work, even when a machine could perform the task.

I am still thinking that one over. This whole category probably deserves a piece of its own.

The line I cannot get comfortable with is the broader observation that governments, companies, and public institutions are poorly equipped for the speed and breadth of the transition. One popular consolation is that companies move so slowly they will blunt the worst effects.

If the only thing preventing a worse outcome is our inability to change how work gets done, then success at transformation removes the safety feature. The same company that finally learns to deploy agents quickly will discover it can also scale a bad incentive, a weak permission model, or a foolish decision much faster than before.

The work ahead is larger than adopting tools. Work, people, governance, and value all have to evolve together. Otherwise we build an AI-using company that moves faster while carrying every old organizational pathology along for the ride.

Two health stories this week were considerably less theatrical and far more encouraging.

Researchers working with Ai2's AutoDiscovery system found a promising cancer signal in data scientists had possessed for years. (If you are a long time listener/reader you know I am a fan of their work.) The system could ask an unreasonable number of questions across large datasets, including questions no research team had found the time or cycles to pursue. Humans then chose which findings deserved attention, checked the signal in an independent dataset, and took it into the lab. It remains a research hypothesis and still needs clinical testing.

That division of work is worth studying. The AI widened the field of possible questions. Scientists supplied judgment, skepticism, and validation. Together they found something neither side had found alone.

The other item is simpler. The FDA authorized a wearable that continuously monitors glucose and ketones for people with diabetes. A second continuous signal can warn that risk is changing before a person would ordinarily stop and run another test. Its effects on diabetic ketoacidosis, adherence, access, and cost still have to be demonstrated in real use.

This is personalized care becoming literal. Your measurements, your direction of travel, your warning at the moment it can still help.

It also extends the shadow healthcare system growing outside the traditional visit. Watches, home sensors, AI assistants, patient communities, and consumer labs are assembling a parallel layer of observation and decision support faster than most health systems can absorb it. That can create earlier intervention and better individual care. It can also create uneven access, anxious over-measurement, and decisions that never make it back into the clinical record.

The answer will emerge from the design of the relationship between the person, the machine, and the clinician. The cancer story gives us a useful shape: machines expand the search, humans decide what survives contact with reality.

There is a great deal happening, and it is happening quickly. NVIDIA is accumulating leverage across the ecosystem. Agents are demonstrating that our oldest management mistakes survive the trip into software. Gates is asking which human choices should be protected while our institutions struggle to move. Health researchers are showing the upside when the collaboration is designed well.

The trouble worth making is the operating environment around all of it. We need systems that distribute capability without surrendering judgment, that reward the goal rather than the dashboard, and that let people move quickly without turning organizational slowness into the last remaining safeguard.

So I will leave you with the uncomfortable version of the question:

If your protection is that your company cannot move fast enough to deploy the capability, what happens on the day it can?

Watch this week's ChuckChat for the live version, the rest of the stories, and a few opinions I failed to keep to myself: https://youtu.be/SWsZULz812E

Chuck
Chief Technology Troublemaker
chuckdevries.ai

Someone forwarded you trouble.
They clearly think you can handle it. Trouble Worth Making is where I work out loud on how to stop bolting AI onto a pre-AI company and actually rebuild the operating model underneath it. One issue every week or two, no filler, the occasional uncomfortable question.

If the person who sent this has good taste (obviously, they do), subscribe here and get the next one straight from the source.