Six startups with checks north of a billion dollars each are not, on close inspection, trying to beat OpenAI and Anthropic to superintelligence on the incumbents’ own timeline. They are structured as if that race either stalls or gets won by someone playing a different game entirely. That distinction matters more than the funding totals, because it tells you what these founders actually believe about where the technology is headed, regardless of what they say in pitch decks.
The six are Safe Superintelligence (SSI), founded by Ilya Sutskever; Thinking Machines Lab, led by Mira Murati; Reflection AI, run by Misha Laskin; Ineffable Intelligence, David Silver’s venture; AMI Labs, Yann LeCun’s project; and Discovery Loop, a Jeff Dean venture in which Google holds both roles at once, putting up founding capital and supplying the compute. Forecasting firm FutureSearch modeled each on compute, capital, senior hiring, and the timing of a first frontier-class model, and published the results on August 4.
The forecast medians tell a consistent story. SSI’s frontier model lands around January 2029 in the median case; Reflection AI, June 2029; Thinking Machines, December 2030. Silver’s Ineffable slides to June 2035, Discovery Loop to June 2036, and LeCun’s AMI Labs all the way to December 2037, with a tail extending past 2050. Those are forecasts, not scheduled release dates, and FutureSearch’s own confidence intervals are wide enough to swallow entire product cycles.
What the underlying strategy reveals is sharper than the dates themselves. A founder who genuinely expects LLMs to keep improving on a smooth curve toward general intelligence would raise as much capital as possible, hire as many senior researchers as the market allows, and buy every gigawatt of compute in reach, because scale is the whole bet. A founder who expects a plateau, and thinks the eventual breakthrough comes from a different architecture or a smaller team applied more precisely, should look completely different on paper. FutureSearch’s own bench shows both patterns sitting side by side.
SSI is the clearest case: roughly $8 billion raised (more than most public tallies show, once Nvidia’s July investment is counted alongside earlier rounds) concentrated on a team FutureSearch pegs at around 50 people, with only two publicly identifiable as former senior staff from a major lab. That is capital and compute stacked against headcount, a bet that a small group with priority access to Nvidia’s next-generation hardware beats an army of researchers. Reflection AI is the mirror image: headcount roughly quadrupled in a year to 230, the fastest hiring pace on the bench, while its compute sits on a month-to-month lease at SpaceX’s Colossus 2 site that either party can cancel with 90 days’ notice. One lab is betting on concentration, the other on talent density and speed to open-weight release. Neither looks like a company racing to out-scale Google.
The plainest evidence of a plateau bet sits with LeCun and Silver, whose labs both reject the transformer-based approach that produced every frontier model to date and both carry the longest forecast timelines on the bench. If either is right that language models alone will not reach general intelligence, a decade-long runway is not a weakness. It is the only honest way to build the thing.
None of this changes the resource math. FutureSearch puts the strongest compute medians here a full factor of ten short of the roughly ten-gigawatt scale OpenAI is building toward with Stargate, and the largest capital medians are a fraction of OpenAI’s most recent funding round. If a straight-line path to transformative AI exists, the incumbents get there first on sheer scale. The neolabs are not contesting that outcome. They are pricing in the possibility that the path bends, and building teams and burn rates that only make sense if it does.
For operators, the tell to watch is not funding announcements. It is whether a lab’s hiring pace, compute footprint, and product cadence match its stated timeline, because a founder who claims a decade-long research bet while burning cash like a 2029 product company is not actually betting against superintelligence at all.
These forecasts and figures come from FutureSearch, published August 4, 2026; as a forecasting outfit projecting future outcomes, its dates, medians, and dollar estimates should be read as its modeled projections rather than confirmed facts.
