In Memphis, a decommissioned appliance factory now houses a few hundred thousand GPUs and a gas turbine plant built to feed them. In Hangzhou, a lab published model weights for free and helped wipe roughly $600 billion off Nvidia’s market value in a single trading session. In London and San Francisco, researchers who spent the 2010s publishing openly now negotiate eight-figure packages to stay quiet. None of that is coincidence. It’s what an AGI competition looks like once it stops being a research problem and becomes an industrial one.
The Contenders, and What Each One Is Actually Good At
There’s no formal league table, but the field thins out fast once you filter for who can afford frontier pretraining runs — reportedly north of $1 billion per attempt — and who can do it more than once.
- OpenAI turned a research lab into a consumer product company and is now trying to become an infrastructure company, with the Stargate buildout as the clearest signal of intent. Its edge is distribution and a willingness to ship before the work feels finished.
- Google DeepMind has the deepest research bench, its own silicon, and a habit of publishing work that rivals then spend a year catching up to. Gemini’s multimodal stack is the most complete of any lab’s.
- Anthropic built a business on the premise that reliability outsells raw capability, and enterprise contracts have largely proved the point.
- Meta bet on open weights to commoditise rivals’ moats, then reorganised around a superintelligence group with effectively unlimited funding and a taste for aggressive recruiting.
- xAI compresses timelines by skipping the usual review cadence and building compute faster than most people thought physically possible.
- DeepSeek, Alibaba’s Qwen team, and Mistral keep demonstrating that a well-run lab on a fraction of the budget can close a two-year gap in months.
That last group stings the incumbents most. Mistral spent two years as Europe’s great frontier hope, went quiet on flagship releases while it built enterprise tooling, and now looks vindicated — the self-styled ‘DeepSeek of the West’ finally has a frontier model worth arguing about.
Three Inputs Decide Almost Everything
Compute
Training compute for frontier models has grown roughly four to five times per year since 2020, according to Epoch AI’s tracking. That curve is why the competition increasingly looks like an energy and construction race. xAI’s Memphis cluster scaled from 100,000 to well over 200,000 GPUs in under a year. Microsoft committed a reported $80 billion to data centres in a single fiscal year. Grid interconnection queues now matter more to release schedules than algorithmic cleverness does.
Talent
The 2024 acquisition dance — Microsoft and Inflection, Google and Character.AI, Amazon and Adept — was talent acquisition dressed as M&A. Since then the numbers have gone feral: individual researchers reportedly offered packages in the $100 million to $300 million range to move. Small labs can’t compete on salary, so they compete on mission, equity, and the freedom to publish. That’s a real advantage, until it isn’t.
Evaluation
Whoever can measure capability fastest ships fastest. Internal eval suites are now among the most guarded assets in any lab, and public benchmarks are treated as lagging indicators. When OpenAI’s internal maths work started leaking out in staged releases — a run of results that kept arriving faster than anyone could verify them — the pattern was clear: capability was being demonstrated selectively, as a negotiating position.
The Benchmarks Keep Losing
Every public test of machine intelligence has a shelf life, and the shelf is getting shorter. Humanity’s Last Exam launched in early 2025 with frontier models scoring in the single digits; within a year, top systems were climbing into the double digits. ARC-AGI went from a flippant ‘easy for humans, impossible for machines’ curiosity to a leaderboard that labs fight over. SWE-bench Verified saturated so quickly that its creators pushed out a harder successor.
This is the quiet structural problem at the heart of the race: every benchmark built to measure AGI progress has eventually been beaten, usually faster than its designers expected. Which means the scoreboard mostly tells you who invested in eval engineering, not who is closest to general intelligence.
Open Weights Reset the Pricing Floor Every Few Months
Meta’s Llama releases, Qwen’s steady drip of models, and DeepSeek’s R1/V3 line have turned open weights into a permanent deflationary pressure on the closed labs. When a free model lands within a few percentage points of a frontier system on most tasks, the commercial argument for paying premium API rates narrows to a handful of use cases.
The closed labs’ counterargument is that open weights trail by six to nine months and always will, because the expensive part is the next pretraining run, not the last one. That’s true so far. It’s also exactly what people said right before the gap closed to three months.
Safety Frameworks Have Become Competitive Instruments
Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework, and Google’s Frontier Safety Framework all serve a genuine function. They also serve a strategic one. A lab that can credibly promise ‘we will not ship a model that fails our evals’ has a product enterprise buyers and regulators can work with. A lab that can’t say that looks reckless by comparison.
Watch how each company handles its first genuinely uncomfortable capability result. Whoever pauses, publishes an honest account, and takes the commercial hit will tell you more about the state of the field than any leaderboard position.
How the Competition Distorts the Timeline
Competition doesn’t just speed things up — it reshapes which problems get solved first. Because there’s a rival one release away, labs optimise for demonstrable capability: agentic coding benchmarks, maths olympiad problems, long-context recall. Slower, less photogenic work like reliability, calibration, and interpretability gets less attention than it deserves.
That skew is why the year-by-year picture is messier than the hype suggests. Plenty genuinely moved in 2025 and 2026; plenty also stalled quietly, and the honest accounting of what actually progressed toward AGI and what didn’t is more useful than any breathless launch post.
Signals Worth Watching Over the Next Twelve Months
- Multi-day autonomous tasks. A model that completes a week of software engineering work without a human rescue is the single cleanest proxy for agentic reliability. Nobody has it yet.
- Cost per million tokens at frontier quality. A 10x drop changes which products become viable overnight and which labs can afford to keep giving models away.
- The open-weight lag. If it stays near six months, the moat is thin. If it stretches past a year, the closed labs have bought themselves real pricing power.
- Power purchase agreements. Grid connections and gas turbine orders predict release schedules better than blog posts do.
- Retention at the top labs. A quiet exodus of senior researchers shows up in model quality about eighteen months later, without fail.
The lab that wins this probably won’t be the one with the best model in any given quarter. It’ll be the one that can keep securing power, keep its researchers from leaving, keep shipping while its rivals argue about the scoreboard, and keep enough goodwill banked to survive the first release that goes badly wrong. That’s a duller competition than the one headlines describe — and a much harder one to win.