Jensen Huang at Fireworks Forge: "We Love Closed Models," but Specialized Open Ones Matter Most

Nvidia CEO Jensen Huang told the Fireworks Forge event in San Francisco on November 3, 2026 that he loves closed models, and that the idea everyone would rely on one kind of general intelligence "is not logical". His main claim is about value, not cost: most of it, he said, sits in specialized intelligence made by adapting an open frontier model to a domain. The cost claims came from others on stage and are vendor figures.

The primary account is The Register's report, written by Thomas Claburn. The event itself is listed on Fireworks' Forge page and agenda, which put Huang in a fireside chat with Fireworks CEO Lin Qiao at 9:00 am at Pier 27. We did not watch the stream, so quotes below are as The Register printed them.

What did Huang actually say?

The "superintelligence" line needs care. Huang's usage is domain-specific: a model adapted to a company's data and tools that outperforms general models on that company's tasks. That is a claim about narrow performance. It is not the sense in which safety researchers use the word.

Two rows of server racks, one open and one closed, linked by a cable, illustrating Nvidia's stance on open and closed models

Why does Nvidia take both sides?

The Register's headline reading is that Nvidia is "happy to take your money" either way. The logic is simple. The cloud customers that serve Anthropic's and OpenAI's proprietary models buy Nvidia GPUs. So do companies that fine-tune and serve open models. Fireworks, which sells a platform to train and run tuned models, "just received an infusion of 25,000 Nvidia GPUs to support its expansion in Europe", per The Register.

Huang defended open weights earlier, in July, against protectionist pressure from the leadership of Anthropic and OpenAI, per the same report. Nvidia has also released its own open models, such as the Nemotron 3.5 family, and expanded its work with Meta on Muse, covered in Nvidia and Meta's Muse collaboration. That makes his dual message consistent with Nvidia's product line, not just a stage remark.

What did the cost claims say, and are they verified?

Fireworks CEO Lin Qiao opened with a slide of ten headlines on runaway token spending. "Cost is out of control," she said, adding that business cannot scale without borrowing a lot of money. A session by Genspark CTO Kay Zhu then presented Gen-1 Slides, a model for generating slide decks.

The ratio is arithmetic, but the comparison is thin. A price per million tokens ignores how many tokens each model uses per slide deck, whether output quality matches on your slides, and which Opus 5 price tier applies. Our existing posts on similar vendor comparisons show the pattern: Shopify's Sidekick post-trained on Nemotron and Samaya Sterling both report large savings on a narrow task, and both are company write-ups.

Independent signals point the same direction but are not proof. Gartner's view that open-weight models handle most agent steps and the Artificial Analysis open-weight gap of 12 points both suggest that open models trail frontier ones on general tasks yet can be good enough on specific ones.

A coin jar with a clamp and a weekly tag, representing the token spending budgets that Forge speakers said are out of control

What else was on the Forge agenda?

The public agenda lists sessions from Meta (post-training and Muse on Fireworks), Microsoft's Jay Parikh, Replit, Cognition, Reflection AI's Misha Laskin, Mercor, Salesforce, Zillow, PayPal, SemiAnalysis's Dylan Patel and Nathan Lambert on how open models win. Themes repeat across titles: post-training, evaluation pipelines, routers and domain-specific models.

That matches the pattern in our other coverage. Lambert's argument that an open-weight Mythos would have been fine is one side of the safety debate. Reflection's open-weights release and the MiniMax M3.1 Flash weights show the supply side. Together's post-training price of $9 per B200 hour shows how inference and training vendors compete on the cost that Qiao highlighted.

How should you test a "frontier quality at a fraction of the cost" claim?

Claims like Genspark's follow a pattern, and a short checklist makes them testable. Run it before you move a workload.

1. Fix the task. Slide decks, fraud flags and support replies each have different failure costs. A cheap model that is wrong 5 percent more often on legal text is not cheap.
2. Collect 50 to 200 real inputs. Use your own data, not the vendor demo set. Include the ugly cases.
3. Score blind. Have two reviewers rate outputs from both models without labels, or use a rubric with a separate judge model that you also spot-check by hand.
4. Count tokens end to end. Record input and output tokens, retries and tool calls per finished task. Cost per task, not cost per token, is the number that matters.
5. Price the whole path. Include fine-tuning or post-training cost, hosting, evaluation upkeep and the engineer time to maintain a custom model.
6. Re-run when models change. A new frontier release can erase the gap, and an open model update can widen it.

Applied to this event: $0.30 against $5 per million tokens is a real gap on paper, but the open question is quality parity on your own decks and the token count each model burns per deck. Huang's own recipe, an open frontier model plus domain knowledge, tools and evaluation, puts evaluation last on the list for a reason: it is the step that tells you whether the first three worked.

What did Huang say about OpenShell and robots?

Asked what might become common next year but seems unlikely today, Huang pointed first to the agent harness, which turned models into agents this year. He then predicted that OpenShell, Nvidia's runtime, would let every company "deploy agents safely, confidently", with the caveat that this is a little self-serving. We covered the project when it hit Hacker News in what developers say about NVIDIA OpenShell and at GTC Berlin.

His second prediction, hardware in the loop, refers to robotic systems closing the loop between AI models and physical hardware. The Register notes Nvidia's evident investment in robotics. Neither prediction came with a date or a metric.

What does this mean for teams choosing models?

1. Do not read Huang's remarks as a verdict on quality. He said both camps have a place. That is a market position, not a benchmark.
2. Test on your own task before believing a price ratio. A 17x per-token gap shrinks if the cheaper model needs more retries or tokens.
3. Budget for evaluation. Huang's recipe includes "the proper evaluation", and every vendor on the agenda sells or discusses eval pipelines for a reason.
4. Treat agent safety as part of the stack. If you adopt OpenShell or a similar runtime, check what it logs and restricts. For agents with write access, AgentBeam from the explainx.ai team is built to stop AI agents before they take dangerous actions.
5. Expect more vendor-reported cost wins. The next claim will likely be a narrow task with a fine-tuned open model. Ask for the eval set, the token counts and the baseline version.

What is verified and what is not

• On the record: Huang's quoted remarks, Qiao's remarks, the 25,000-GPU infusion for Fireworks' European expansion, and the event agenda.
• Vendor claims: Genspark's quality parity with Opus 5 and its price comparison.
• Opinion: "superintelligence is already here", and the two predictions.
• Not found: any independent benchmark of Gen-1 Slides in the coverage we read, or any Nvidia pricing statement.

Related reading

• Nvidia and Meta expand the Muse collaboration
• Shopify Sidekick on Nemotron, claimed vs verified
• Gartner and open-weight models in agents
• NVIDIA OpenShell on Hacker News
• MiniMax M3.1 Flash weights
• Artificial Analysis open-weight gap

Quotes reflect The Register's November 3, 2026 report and may change as the Forge recording is reviewed.