SiMa.ai Hits $1.45B on Edge Bet; AMD Buys World Labs

SiMa.ai closed $150 million at a $1.45 billion valuation this week, led by Fidelity and Amplify Partners, to build edge inference chips that don't need a data center or an Nvidia sticker. The company's pitch: physical AI—robots, drones, cars, security cams—can't wait for cloud latency or afford the power budget. Its MLSoC platform runs vision models locally at under 5 watts. Customers include unnamed automotive OEMs and a robotics deployment pipeline that allegedly includes humanoids. No revenue disclosed, no ship dates confirmed, but the investor list now includes Point72, Jericho, and Maverick, so someone believes the math.

The timing matters. The same 48 hours saw AMD acquire World Labs, the 3D scene-understanding startup founded by AI researcher Fei-Fei Li, for an undisclosed sum. World Labs builds models that turn 2D images into navigable 3D environments—core infrastructure for robots that need to understand a room without crashing into it. AMD has been chasing Nvidia in AI accelerators for two years; this is the first acquisition that says 'robotics' out loud. No integration timeline, no product roadmap, just a signal that the edge is now a battlefield.

Meanwhile, Singapore's Home Team announced humanoid robots will deploy alongside officers by 2028. No vendor named, no unit count, no budget. Horizon Robotics CEO Yu Kai told investors that whoever wins autonomous EVs will win robotics, projecting 'mind-off driving' by 2035 and claiming the sensor stack is 80 percent identical. SemiDrive and Radxa launched the RoboX D9M motion controller for under $200, targeting hobbyists and integrators who need compute without the Jetson Orin price tag.

Back in the cloud, Goldman Sachs published new math: the Mag 7 have a combined $1.69 trillion AI capex backlog and need to generate $1 trillion in incremental revenue to hit a 15 percent return. Micron reports earnings tonight; consensus expects net income to jump 10x year-over-year on HBM demand, but analysts are watching guidance for signs the capex cycle is peaking. Bain put out two reports in two days arguing AI needs $6 trillion in total investment to justify current buildout—roughly 4x today's spend—and that 'new innovation' in financing is required because traditional equity and debt won't cover it.

KBW noted that U.S. banks' lending commitments to data center and AI infrastructure projects are up, signaling confidence the cycle has legs. The counterpoint: one economist told Fortune that AI agents negotiating contracts and procurement could prove deflationary, compressing margins for everyone who thought they'd ride the capex wave forever. Southeast Asia is now seeing its own data center boom, with several hundred megawatts of capacity planned across Vietnam, Thailand, and Indonesia, though power grid constraints remain unaddressed.

Nvidia alumni are spinning out at record pace—Business Insider counted at least a dozen new robotics and AI hardware startups in Q3 alone, most focused on edge inference, vision models, or motion control. None have shipped at scale. SiMa.ai is the first to crack $1 billion in paper value without shipping a consumer product, which either means edge inference is about to explode or venture markups have lost all tether to reality.

One headline this week: 'Gone rogue and set to kill us all? Please. I know the banal truth.' The banal truth is that most robots still can't climb stairs, most edge chips are still in pilot programs, and most of the $1.69 trillion backlog is still concrete and copper, not code. But Fidelity doesn't usually write checks for chips that run on drones, and AMD doesn't usually buy 3D vision labs. Something is moving.