Competition

The tell is in what they stopped promising. For two years Anthropic sold itself as the lab that would slam the brakes if safety couldn't keep pace, the central pillar of its Responsible Scaling Policy. In late February 2026 it quietly scrapped that pledge. The reason was not new science. It was the scoreboard. "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments... if competitors are blazing ahead," chief scientist Jared Kaplan told TIME (Feb 24, 2026). That sentence is the whole debate compressed. Everyone in this field agrees the race is dangerous. Nobody will be the one to slow down. Competition, for the frontier labs, is the gravitational field that bends every other principle.

The race eats the guardrails

Watch the mechanism. Anthropic's old RSP categorically barred training models past a capability threshold unless mitigations were in place. The new version "commits to matching or surpassing the safety efforts of competitors" and promises to delay development only if leaders both believe Anthropic is leading the race and think catastrophe is likely (TIME, Feb 24, 2026). Safety became relative, indexed to the field rather than to any fixed line. Kaplan denied this was capitulation: stopping would just make Anthropic "lose relevance as an innovator who understands the frontier." METR's Chris Painter, who reviewed an early draft, called it understandable but "a bearish signal," warning the move away from hard tripwires invites a "frog-boiling effect, where danger slowly ramps up without a single moment that sets off alarms."

The same logic drives the parts that look like spending, not retreat. Dario Amodei told the Dwarkesh Podcast he spends "a third, maybe 40%" of his time on company culture, because at a firm now valued at $380 billion with 2,500 employees, holding everyone "toward the same mission instead of against one another, as he said happens at other unnamed AI companies," is "the only thing that will win the AI race" (via Fortune, Feb 26, 2026). Even culture is a competitive weapon. Greg Brockman's version is more literal: the binding constraint is physical. "We are heading to a world of absolute compute scarcity," he told Matthew Berman (Oct 8, 2025), with energy the looming bottleneck in the US. OpenAI now builds its own data centers (Stargate), spreads across NVIDIA and AMD, and treats compute as "this fundamental reagent." The race isn't an idea you win with cleverness. It's a buildout.

The fault line: collaborate or capture

Here the leaders genuinely split. Ilya Sutskever's whole pitch is that competition will eventually dissolve into cooperation. At TED (Nov 20, 2023) he argued people will "start to act in unprecedentedly collaborative way out of their own self-interest" as the stakes become obvious, citing the Frontier Model Forum and OpenAI's longstanding line that if a rival got close to AGI first, "rather than compete with them, we will help them out, join them." His company is named for the escape hatch: Safe Superintelligence, built to "scale in peace," with a business model deliberately "insulated from short-term commercial pressures" (SSI launch, Jun 21, 2024). It is the purest bet that you can step outside the race by refusing to ship.

Dario's worry runs the other way. On People by WTF (Feb 25, 2026) he said he is "at least somewhat uncomfortable with the amount of concentration of power that's happening here... almost overnight, almost by accident," pointing to Anthropic's Long Term Benefit Trust as a hedge. So you get the contradiction at the center of Anthropic: a founder who built unusual governance because he fears winners taking everything, running a company that just loosened its safety brakes so it could stay in the running. Sutskever thinks the danger summons cooperation. Amodei thinks it summons concentration. Both are watching the same race; they disagree about what it ends in.

When the rivalry turned into a knife fight

For most of 2025 the competition was civilized, fought with valuations and benchmarks. Then in late February 2026 it got operational. The Pentagon demanded Anthropic let the military use Claude "in all lawful use cases without limitation"; Anthropic refused over autonomous weapons and domestic mass surveillance, and the DOD moved to brand it a "supply chain risk." What happened next is the most revealing competitive episode in the file. Some 70 OpenAI staffers signed an open letter, "We Will Not Be Divided," in support of a rival. Sam Altman wrote his own staff that "this is no longer just an issue between Anthropic and the Pentagon; this is an issue for the whole industry," declaring OpenAI shared Anthropic's red lines (Axios/CNBC, Feb 27, 2026).

Solidarity, right up until the deal. By Friday night OpenAI had struck its own agreement with the Pentagon while Anthropic announced it would sue the department after being blacklisted (Axios, Feb 27, 2026). Altman could plausibly have OpenAI's models replace Claude in classified work. The labs found a shared red line and a shared adversary, and still the competitive logic delivered exactly what you'd expect: one signs, one litigates, market share changes hands. The "We Will Not Be Divided" letter aged in about 36 hours.

The other rivalry is over who counts as a peer

Then there's China, and here the frontier leaders are oddly aligned on a comforting story. Demis Hassabis (AI Nutshell, Feb 16, 2026) granted that Chinese labs are now "maybe only a matter of months behind" the Western frontier, with DeepSeek's January 2025 breakout and Alibaba's Qwen among the most-downloaded models anywhere. But he drew a hard line: "to invent something is about 100 times harder than it is to copy," and he hasn't seen evidence China can produce something genuinely new "like a new transformer." He framed it not as a chip-access problem but a "mentality issue," casting DeepMind as a modern Bell Labs that prizes "exploratory innovation, not just scaling out what's known." Replication versus invention is the moat he's betting on. It is also, conveniently, the one moat export controls can't hand you and incumbency can.

Even the friendly rivalry runs on the same fuel. Altman, asked in India (Indian Express, Feb 20, 2026) what he admires about Google's catch-up, praised Demis for working on AI "before anyone else... with a lot of conviction" and then Google's "relentless focus and execution... after being pretty far behind." Generous, and also a reminder that the man crediting his rival is the one who put ChatGPT out and forced everyone, Google included, into the sprint.

Where the competition has no script

The quieter fault line is what they're competing for on the way down, and here Daniela Amodei drew a real product distinction. Anthropic's first Super Bowl ad vowed never to put advertising inside Claude, because ad incentives reward keeping "the customer's eyeballs on the model for a longer period of time," which makes resisting sycophancy "much harder if you're in an advertising-based business" (ABC News, Feb 7, 2026). Altman fired back on X: "We would obviously never run ads in the way Anthropic depicts them. We are not stupid." Different theories of the consumer business, dressed as a moral fight.

And Jack Clark, on Conversations with Tyler (Feb 26, 2026), pointed at the part nobody is racing to solve: the uneven, messy aftermath. He predicted "a high chance for a political movement... which tries to freeze a load of human jobs in bureaucratic amber," driven not by reason but "by the chaotic winds of political forces." The labs compete furiously over who builds the technology. On who absorbs the shock, the field goes quiet.

So the consensus holds and means nothing. Every leader will tell you the race is reckless. Then Kaplan rewrites the pledge, Altman signs the contract, Hassabis guards the moat, and the whole industry keeps running, each one certain that the only thing more dangerous than the race is losing it.