Robotics: capex curve is finally bending toward physical AI  April 25th 2026

The Big Story

Tesla raises 2026 capex to $25B for AI, Optimus production, and robotaxi build-out · April 22 · [CNBC]
→ Tesla bumped 2026 guidance from $20B to $25B in capex, earmarked for AI training, Optimus humanoid production starting July, robotaxi factories, and an Austin chip fab. CFO confirmed Tesla will go free-cash-flow-negative for the rest of 2026. The capex curve and the org chart now both bend toward physical AI on a multi-year horizon — and the company is pricing its self-funding model accordingly.

Also This Week

Figure AI now builds a humanoid every 90 minutes · April 22 · [RoboHorizon]
→ BotQ's first line is rated for 12,000 units/yr; Figure 03 is the first model engineered ground-up for high-volume manufacturing rather than R&D, with BMW and Brookfield commercial agreements already in place.

Waymo opens Nashville, partners with Lyft, eyes 1M weekly rides by year-end · April 7 / running · [TechCrunch]
→ Eleventh public city, but the Lyft tie-up is the real signal — Waymo no longer needs to own demand-side UX to scale, freeing capital for the Magna-built Zeekr Ojai fleet doubling out of Arizona.

Pony.ai cuts Gen-7 robotaxi below RMB 230k, debuts world's first L4 light truck with CATL · April 24 · [PR Newswire]
→ The autonomy-kit-included sticker is now competitive with mid-trim ICE sedans in China — the unit-economics argument against L4 is closing faster than the regulatory one.

Symbotic acquires Fox Robotics to fold autonomous forklifts into the platform · March / Q1 print · [Robotics & Automation News]
→ Backlog at $22.3B, Walmart now equity-aligned after the $200M / $520M swap, dock-to-shelf coverage in one stack — the pure-play warehouse-AI story the public markets have been waiting for.

Medtronic Stealth AXiS Autopilot — first U.S. surgeons use AI-enabled robotic platform · April · [MassDevice]
→ Spine and cranial/ENT clearances stacked back-to-back in Q1; the AI layer here isn't autonomy, it's planning + intra-op visualisation, and it's the wedge by which Medtronic finally challenges Intuitive's da Vinci moat.

From the Lab

Sony AI's Ace beats elite human players in table tennis — Nature cover, April 23 · [Nature]
→ Event-based vision tracking the ball at 200 Hz with 3 mm error, fused with deep RL and agile actuation. Won 3-of-5 matches against elite amateurs and beat all three pros at least once. The interesting bit isn't the win column — it's that the perception/control loop is now tight enough to handle adversarial, sub-100ms decision physics outside a lab. That's the same loop you need for surgical assist and mid-air drone interception.

VLM2VLA: turning vision-language models into robot-control policies without forgetting · [Princeton AI Lab / arXiv]
→ Princeton's pipeline fine-tunes VLMs into vision-language-action policies while preserving the original VQA and reasoning capacity — generalisation on real manipulation tasks goes up, not down. This directly attacks the catastrophic-forgetting tax that's been the dirty secret of every "general-purpose" robot policy demo since RT-2.

Worth Reading

• [Humanoid Robots in Industrial Manufacturing: What They Can (and Can't) Do in 2026] — A blunt buyer-side audit of where humanoids actually clear payback today (tote-destacking, kitting) vs. where the marketing decks are still ahead of the joints.
• [Siemens + Humanoid Ltd. deploy HMND 01 Alpha at Erlangen factory] — 60 container moves/hour and >90% pick success in live logistics — the first European humanoid trial with a credible throughput number attached.
• [How Physical AI Is Reshaping Robotics — and What Comes Next] — BCG's framing of the world-model + sim + foundation-policy stack, useful if you're building the buyer-side narrative for a board.

The next twelve months won't be decided by who has the smartest policy. They'll be decided by who can ship the most arms, on the most lines, with the cleanest data flywheel pointed back at the model. Welcome to the BOM era.

— Alexis