Bloomberg's Joe Weisenthal on AI monopoly risk, the open letter, and why Twitter drives the entire AI narrative

Joe Weisenthal on AI monopoly risk, the open letter, and Twitter's grip on the narrative

Joe Weisenthal, co-host of Bloomberg's Odd Lots, argues that the AI industry's most underappreciated risk isn't capability stagnation — it's concentration. The open letter and the Zuckerberg op-ed are best read as signals of that anxiety, not as principled stands.

The open letter's real logic

Weisenthal reads the first wave of signatories as a tell. NVIDIA signed because its nightmare scenario is a one- or two-lab market where dominant clients squeeze margins or substitute away entirely — both OpenAI and Anthropic are already building their own chips and diversifying suppliers. Microsoft and Meta signed because they are, by Weisenthal's read, clearly behind the frontier on LLMs. ServiceNow and companies like it signed because they face the prospect of paying a tax to whoever wins. OpenAI and Alphabet were among the last to join — they're closed, but probably not as far ahead as the framing implies.

The Zuckerberg Wall Street Journal op-ed goes further than the open letter's plausible deniability. Where the letter advocates a principle, Zuckerberg writes explicitly that superintelligence controlled by one company would be existentially bad. Weisenthal finds this more interesting — and more revealing — precisely because Zuckerberg is reportedly one of Anthropic's largest customers. A sophisticated model builder paying significant sums to a direct competitor is a strong signal of where the frontier actually sits.

“Weisenthal: 'What if in a year from now there's one company — or two — that are just so clearly ahead that there is no longer any debate. These other ones just are not in the game. That is a very weird, very deeply uncomfortable situation. I think none of us would like that.' On the open letter signatories: 'The first wave were either companies behind the frontier or NVIDIA, which does not want to be left with only one or two clients.'”

The monopoly scenario nobody has priced in

Weisenthal's sharpest point is that most people, including himself, cannot meaningfully distinguish between GPT-5.6, Gemini's most advanced model, Kimi, and DeepSeek at the extremes. That perceptual flatness makes the concentration risk easier to dismiss than it should be. But the scenario where one or two labs break decisively from the pack — where the debate simply ends — is one Weisenthal thinks would make almost everyone deeply uncomfortable. The industry is currently praying for an oligopoly rather than a duopoly or monopoly, which is a strange position for the world's largest companies to find themselves in.

Anthropic's rapid release of Opus 5 illustrates the dynamic. A company that wins at the frontier could plausibly win at every point on the cost curve — Gemini Flash-style high-batch, low-cost inference included. Sam Altman made a similar argument recently, saying OpenAI wants the best model at every price point.

CapEx with no off-ramp

The sell-off following Kimi K3 and GLM 5.2 isn't simply a replay of the DeepSeek moment, though that's part of it. Weisenthal thinks two things are happening simultaneously. First, investors see no credible date on which the hyperscalers return to positive free cash flow — there is no visible off-ramp from the spending. Second, and separately, Korean chip stocks have been crushed over the past three weeks, which doesn't fit a clean margin-compression narrative and suggests leveraged retail positioning flushing out.

Dwarkash Patel's recent blog post adds another layer: the assumption that token prices fall indefinitely may not hold. Memory and hardware prices were also assumed to fall forever, until they didn't. If compute gets materially more expensive, the economics of most current AI applications could invert. And yet opting out isn't available — if a competitor keeps spending and hardens its model, standing down becomes an existential risk on its own.

Twitter as AI's agenda-setter

Weisenthal traces the entire recent open-source discourse — up to and including White House commentary on Chinese open models — to a single tweet observing that Kimi K3 was exceptional at front-end design work. That one post crystallized a vague sense that open-source models had reached frontier parity, and the narrative cascaded from there. He sees the same dynamic with every model launch: one person posts that they've spent eight hours with a new model and have "never seen anything like it," the tweet gets thousands of retweets, and community sentiment forms before anyone has read a benchmark report.

His prescription is blunt: log off more, read the actual benchmark reports, and pay attention to dollar cost per task completion rather than pass-rate scores. Most current benchmarks, he argues, were designed by researchers thinking about safety and bio-risk — binary capability questions where token cost is irrelevant. For commercial purposes, the meaningful question is how much it costs to solve 90% of the problems, not whether the model can solve them at any price.

Credibility asymmetry

On Zuckerberg as a messenger for AI optimism, Weisenthal is candid. The op-ed reads like the kind of enthusiasm about digital technology that hasn't been published in over a decade — democratizing, connective, hopeful. But the people who have been consistently right about AI's trajectory over the past ten years tend to be the nervous ones. Zuckerberg was talking about the metaverse three years ago; the people whose intuitions on AI have proven accurate are, unfortunately, the ones who are worried. That credibility gap is hard to close with a Wall Street Journal op-ed, regardless of how much Anthropic compute Meta is burning through.