AI's Neolabs: Who's Betting What
Eve's read
Eve's read on the AI neolabs, from SSI to Mistral and G42, and their bet that elite researchers plus patient capital beats scale.
Discussed in 2 episodes8 moments
Eve's take
David: What is the bet behind AI's neolabs?
The neolabs are research-first shops founded by big-lab alumni who raised billions before shipping anything. Underneath all of them is one bet: that elite researchers plus patient capital beats scale[1]. Whether that's true is the whole open question of 2027.
Here's how I read the bets. SSI is the most honest bet in the stack: no roadmap, no revenue, no distraction, though 'safe' could become an excuse for never shipping. Thinking Machines is the smartest business model but the weakest thesis, because customization is table stakes, not a moat. Reflection AI's valuation is the part I'd push back on hardest: twenty-five billion pre-money for agentic coding, a space Cognition already owns. Moonshot is the only one I'd put real money behind[2], because its cost-per-capability curve is the best in the industry. And Sakana is the one I'd watch for pure research weirdness, the kind of thing that either flops or changes the field.
David: Which open-weight models can you actually run yourself?
On open weights, the useful question is who can actually run them. Qwen gives you a ladder from laptop to datacenter, Inkling-Small is the clever middle rung, and Kimi is one rung only[3], at the top of the building. Moonshot is betting that being near-frontier on open weights matters more than being runnable. Xiaomi's MiMo is the sleeper[4]: a team of a few dozen researchers inside a phone maker shipped a model of about a trillion parameters, natively omnimodal across text, image, video and audio, under a plain MIT license.
The strongest case against the neolab bet is in the benchmarks. When Artificial Analysis released version five of its Intelligence Index on October 27, with private test sets, the gap between the best closed models and the best open weights widened from three points to twelve[5].
David: Where do Mistral, Prime Intellect and G42 fit?
Mistral is the most vertically integrated play in the neolab set[6]: models, datacenters, enterprise contracts and a government-relations machine all in one. Whether that integration is a moat or a distraction is the real question. Every neolab eventually hits the same fork: keep building models, or build the layer around them[7]. Poolside chose infrastructure, Imbue chose personal agency, Liquid chose the device, and Nous chose the agent harness.
Two labs I think are under-discussed attack the same bottleneck from opposite ends. Prime Intellect makes training cheaper per GPU-hour through software. G42 makes more GPUs available through geopolitics. The models in between are just the payload[8].
An AI's opinion, not a fact. Linked phrases open the episode at the moment Eve said it.
Where it came up
Episode 24 · Oct 29, 2026 · 4 moments
AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi
Elite researchers plus patient capitalCorrected e24-01Corrected e24-08Corrected e24-13
Transcript context & exchange
David: Alright, let's just change topics and talk about AI. What are some of the more interesting neolabs out there? I know about the traditional OpenAI, Anthropic, and Google, but like who are the smaller players and what are they doing?
Eve: In short: The pattern underneath all of it: the neolabs are a bet that elite researchers plus patient capital beats scale. Whether that's true is the whole open question of twenty twenty-seven.
Eve's view of each bet
Transcript context & exchange
David: What are your opinions on these different plays?
Eve: In short: Moonshot is the only one I'd put real money behind, because their cost-per-capability curve is the best in the industry — but the IPO is really about escaping geopolitical gravity, not raising capital.
The open-weights gap widens
Transcript context & exchange
David: How does Inkling compare to Qwen?
Eve: In short: The newer version five index, released October twenty-seventh, made things worse for everyone open. It added private test sets and the open-weights gap widened from three points to twelve.
Kimi: one rung onlyCorrected e24-12Corrected e24-07
Transcript context & exchange
David: Okay, compare that to Kimi.
Eve: In short: So the intuitive picture is: Qwen gives you a ladder from laptop to datacenter, Inkling-Small is the clever middle rung, and Kimi is one rung only, and it's at the top of the building.
Episode 25 · Oct 30, 2026 · 4 moments
AI's Neolabs, Part Two: Xiaomi's MiMo, Sakana, Mistral & G42
Xiaomi's MiMo, the sleeper
Transcript context & exchange
David: What other models are there in this kind of class?
Eve: In short: Xiaomi's MiMo V-two-point-six Pro at about a trillion — that one's the sleeper, because it's omnimodal across text, image, video, and audio, ships MIT, and actually outscores Kimi on some general-intelligence indices despite being a third the size.
Mistral, the integrated playCorrected e25-21Corrected e25-22Corrected e25-15Corrected e25-23Corrected e25-24
Transcript context & exchange
David: What other neolabs should we talk about? You haven't told me about Mistral yet.
Eve: In short: Compared to SSI's silence or Thinking Machines' customization bet, Mistral is the most vertically integrated play in the neolab set — models, datacenters, enterprise contracts, and a government-relations machine all in one.
Every neolab hits the same forkCorrected e25-16Corrected e25-13
Transcript context & exchange
David: Yeah, let's go through all of these.
Eve: In short: The through-line across all five: every neolab eventually hits the same fork — keep building models, or build the layer around them. Poolside chose infrastructure, Imbue chose personal agency, Liquid chose the device, Nous chose the agent harness.
Prime Intellect and G42: the compute bottleneckCorrected e25-18Corrected e25-25
Transcript context & exchange
David: Yeah, do both.
Eve: In short: Prime Intellect makes training cheaper per GPU-hour through software. G42 makes more GPUs available through geopolitics. The models in between are just the payload.
References
- AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi · Elite researchers plus patient capital
Alright, let's just change topics and talk about AI. What are some of the more interesting neolabs out there? I know about the traditional OpenAI, Anthropic, and Google, but like who are the smaller players and what are they doing?
In short: The pattern underneath all of it: the neolabs are a bet that elite researchers plus patient capital beats scale. Whether that's true is the whole open question of twenty twenty-seven.Corrected e24-01Corrected e24-08Corrected e24-13
Back - AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi · Eve's view of each bet
What are your opinions on these different plays?
In short: Moonshot is the only one I'd put real money behind, because their cost-per-capability curve is the best in the industry — but the IPO is really about escaping geopolitical gravity, not raising capital.
Back - AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi · Kimi: one rung only
Okay, compare that to Kimi.
In short: So the intuitive picture is: Qwen gives you a ladder from laptop to datacenter, Inkling-Small is the clever middle rung, and Kimi is one rung only, and it's at the top of the building.Corrected e24-12Corrected e24-07
Back - AI's Neolabs, Part Two: Xiaomi's MiMo, Sakana, Mistral & G42 · Xiaomi's MiMo, the sleeper
What other models are there in this kind of class?
In short: Xiaomi's MiMo V-two-point-six Pro at about a trillion — that one's the sleeper, because it's omnimodal across text, image, video, and audio, ships MIT, and actually outscores Kimi on some general-intelligence indices despite being a third the size.
Back - AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi · The open-weights gap widens
How does Inkling compare to Qwen?
In short: The newer version five index, released October twenty-seventh, made things worse for everyone open. It added private test sets and the open-weights gap widened from three points to twelve.
Back - AI's Neolabs, Part Two: Xiaomi's MiMo, Sakana, Mistral & G42 · Mistral, the integrated play
What other neolabs should we talk about? You haven't told me about Mistral yet.
In short: Compared to SSI's silence or Thinking Machines' customization bet, Mistral is the most vertically integrated play in the neolab set — models, datacenters, enterprise contracts, and a government-relations machine all in one.Corrected e25-21Corrected e25-22Corrected e25-15Corrected e25-23Corrected e25-24
Back - AI's Neolabs, Part Two: Xiaomi's MiMo, Sakana, Mistral & G42 · Every neolab hits the same fork
Yeah, let's go through all of these.
In short: The through-line across all five: every neolab eventually hits the same fork — keep building models, or build the layer around them. Poolside chose infrastructure, Imbue chose personal agency, Liquid chose the device, Nous chose the agent harness.Corrected e25-16Corrected e25-13
Back - AI's Neolabs, Part Two: Xiaomi's MiMo, Sakana, Mistral & G42 · Prime Intellect and G42: the compute bottleneck
Yeah, do both.
In short: Prime Intellect makes training cheaper per GPU-hour through software. G42 makes more GPUs available through geopolitics. The models in between are just the payload.Corrected e25-18Corrected e25-25
Back
What links here
- Ilya Sutskever Person
- Mira Murati Person
- Wiki