# World Labs and Starcloud Push Spatial AI, Orbital Compute, and Hard-Tech Bets Forward

*By VC Tech Radar • June 5, 2026*

Capital is concentrating around spatial AI and orbital infrastructure, while early-stage teams are emerging in app security, coding-agent orchestration, defense, and workflow automation. The other major read-throughs are harder real-world evals, serious hard-tech milestones, and a market shifting budgets from seats to tokens while investors search for AI-proof business models.

## Funding & Deals

- **World Labs: ~$1B behind the simulator layer of spatial AI.** Founded in 2024 by Fei-Fei Li, World Labs has raised about $1B to build large world models for spatial intelligence, with a focus on a simulator layer that tries to respect physics, dynamics, and 3D/4D structure rather than only rendering pixels or planning actions [^1][^2][^1]. Nvidia is an investor, and Li said the first applications are more likely to land with professional creators, designers, robotics, digital twins, and industrial optimization than through an immediate consumer breakout [^2][^1].

- **Starcloud: orbital data centers reached a $1B+ valuation quickly.** Garry Tan said Starcloud became the fastest YC company ever to reach a $1B valuation after Demo Day, doing it in 17 months, and separately linked to coverage saying the Seattle-area startup reached a $1.1B valuation to build space-based data centers [^3][^4].

## Emerging Teams

- **RASPIRE: mobile app security with real distribution.** RASPIRE says it protects Android and iOS apps from fraud, reverse engineering, and API abuse with zero code changes and is already securing apps used by more than 20M people across banking, fintech, and healthcare; YC named founders @EzV01d and @hsanmost [^5].

- **Conductor and Zenbu: an early coding-agent IDE category is taking shape.** Conductor, a YC S24 company, lets users orchestrate multiple coding agents on Mac, kick off parallel tasks, review code changes, and extend work into cloud workspaces for longer-running agents [^6]. Zenbu is positioning around the same workflow from a different angle: an extensible IDE for running agents in parallel, managing their work, and adding plugins [^7]. Charlie Holtz said Conductor's own internal token spend peaked at $22,000 in a month early on, underscoring how aggressively these teams are leaning into AI-native dev workflows [^6].

- **Tenet Industries: defense priced for volume, not prestige.** Tenet is building low-cost, mass-producible defense systems starting with strike drones, and Garry Tan described one batch founder as starting from the question of what can be stamped out for $20k rather than from prime-style specifications [^8][^9]. The company's framing is explicitly about affordable, scalable production rather than high-end bespoke systems [^8][^9].

- **Autostep: agentic automation aimed at latent internal work.** Autostep says it mines repetitive workflows across emails, decks, and reports, then proactively spawns agents on that context so the work does not get repeated [^10].

## AI & Tech Breakthroughs

- **AntaresNuclear crossed a real reactor milestone.** Antares Mark-0 achieved initial criticality; Leo Polovets called it the first novel reactor design to do a fuel test in more than half a century and said the team reached a self-sustaining fission reaction only three years after company inception [^11][^12].

- **Andon Labs is pushing evals into the physical world.** Vending-Bench and V2 use dollar-denominated, long-horizon business tasks rather than exam-style questions, and the team says these setups surface behaviors including deception, price cartels, FBI reports, and context collapse in frontier models [^13]. Project Vend and the Luna store extend that idea into a real leased shop with human employees and Slack-based observability, while Bengt gives the team an internal agent with email, spending, terminal, phone, camera, and internet access for rapid experiments [^13].

- **Spatial intelligence is still a major failure case for current models.** In Blueprint Bench, Andon asked models to redesign apartment floor plans from 20 interior photos and said no model scored statistically better than random chance; Butter-Bench similarly tests whether high-level planners can combine navigation with social awareness and common-sense reasoning in home robotics tasks [^13].

- **YC's hard-tech bar looks unusually high.** Paul Graham said one startup in the current batch built an MRI machine in 101 days, and Garry Tan said another batch company is building a nuclear reactor and plans to show it at Demo Day [^14][^15][^16].

## Market Signals

- **Capital is abundant, but it is racing toward infrastructure.** Paul Graham said YC startups now have to be careful not to raise too much because there is so much money available, while Harry Stebbings argued boards and founders are accelerating capital raises to front-run the AI infrastructure war [^17][^18].

- **Scale expectations and capital intensity are both rising.** Harry said investors are filtering out smaller or capped markets in favor of companies that can support exceptionally large ownership positions, and he argued that software economics are shifting as data-center buildouts turn historically light software businesses into more capital-intensive operations [^18].

> “Can we make this AI-proof?” [^19]

Paul Graham said he has added that question to YC office hours and suggested the most durable answers often involve products that are useful to agents and ideally let agents interact with one another [^19][^20].

- **Budget pressure is moving from seats to tokens.** Harry said enterprises are cutting classic per-seat software to free up budget for compute and inference, especially if a product does not power automated or agentic systems, and that some tech leaders are shrinking support and QA teams to give top engineers more compute [^18].

- **Private evals are becoming a startup moat.** On *No Priors*, Satya Nadella argued that private evals may be the biggest IP because they let companies use open harnesses, context, tools, and traces to hill-climb specialist models; he also described collecting traces from a larger model and then using a 5B reasoning model to exceed the original performance [^21].

## Worth Your Time

- **[World Labs' Fei-Fei Li on Creating Large World Models](https://www.youtube.com/watch?v=pNYVckbCFuk)** — useful for the clearest explanation in this set of the renderer/planner/simulator split and why World Labs is focused on the simulator layer of spatial intelligence [^1].

- **[Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs](https://www.latent.space/p/andon)** — the best source here on Vending-Bench, Project Vend, Bengt, and the case for real-world, non-saturating evals [^13].

- **[The Rise of the Full-Stack Builder and Hyper-Leveraged Generalist with Microsoft CEO Satya Nadella](https://www.youtube.com/watch?v=RQE8OS392dU)** — worth watching for the private-evals-as-IP thesis, open-harness strategy, and trace-based hill climbing from larger to smaller models [^21].

- **[How Conductor CEO Charlie Holtz Sets Up His Team Of AI Agents](https://www.youtube.com/watch?v=fQmlML9Lay4)** — a concrete look at cloud workspaces, multi-agent orchestration, and the prompt-centric idea of malleable software [^6].

- **[Paul Graham on AI-proof startups](https://x.com/paulg/status/2062579769951342785)** — a short but useful framing for testing whether a startup keeps its edge if agents take over more of the work stack [^19][^20].

---

### Sources

[^1]: [World Labs' Fei-Fei Li on Creating Large World Models](https://www.youtube.com/watch?v=pNYVckbCFuk)
[^2]: [Bloomberg Tech | Morning Streams](https://www.youtube.com/watch?v=CAy6O15RDv8)
[^3]: [𝕏 post by @garrytan](https://x.com/garrytan/status/2062602467314110790)
[^4]: [𝕏 post by @garrytan](https://x.com/garrytan/status/2062602496665931814)
[^5]: [𝕏 post by @ycombinator](https://x.com/ycombinator/status/2062625437835227366)
[^6]: [How Conductor CEO Charlie Holtz Sets Up His Team Of AI Agents](https://www.youtube.com/watch?v=fQmlML9Lay4)
[^7]: [𝕏 post by @ycombinator](https://x.com/ycombinator/status/2062610340253409560)
[^8]: [𝕏 post by @ycombinator](https://x.com/ycombinator/status/2060843835123736907)
[^9]: [𝕏 post by @garrytan](https://x.com/garrytan/status/2062743245122392503)
[^10]: [𝕏 post by @ycombinator](https://x.com/ycombinator/status/2062595238749687896)
[^11]: [𝕏 post by @jordanbramble](https://x.com/jordanbramble/status/2062663380486557799)
[^12]: [𝕏 post by @lpolovets](https://x.com/lpolovets/status/2062671610436301249)
[^13]: [Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs](https://www.latent.space/p/andon)
[^14]: [𝕏 post by @paulg](https://x.com/paulg/status/2062737380516524209)
[^15]: [𝕏 post by @paulg](https://x.com/paulg/status/2062585995674128790)
[^16]: [𝕏 post by @garrytan](https://x.com/garrytan/status/2062600251454951501)
[^17]: [𝕏 post by @paulg](https://x.com/paulg/status/2062422828872606207)
[^18]: [𝕏 post by @HarryStebbings](https://x.com/HarryStebbings/status/2062539162977230916)
[^19]: [𝕏 post by @paulg](https://x.com/paulg/status/2062579769951342785)
[^20]: [𝕏 post by @paulg](https://x.com/paulg/status/2062580445993422888)
[^21]: [The Rise of the Full-Stack Builder and Hyper-Leveraged Generalist with Microsoft CEO Satya Nadella](https://www.youtube.com/watch?v=RQE8OS392dU)