AI Daily — Apr 30, 2026

The robots aren't coming — they're already clocking in. Figure AI says it's now building one humanoid per hour. Japan Airlines is putting Chinese-made humanoids on the Haneda tarmac in May. And Meta just told investors it plans to spend up to $145 billion on AI infrastructure this year, then watched its stock fall 6% in afterhours trading on April 29, 2026. Today is the day "AI hype" met its three hardest constraints: factories, airports, and balance sheets.

• Claude Opus 4.7 (Anthropic): Top-ranked frontier model on agentic tasks per community evals, reportedly ahead of GPT-5.4 and Gemini 3.1 Pro.
• Kimi K2.6 (Moonshot AI): 256K context window, now #1 on OpenRouter's weekly leaderboard.
• NVIDIA Nemotron 3 Nano Omni (NVIDIA): Open 30B multimodal MoE with 256K context, English-only speech via a Parakeet encoder, 5.95% WER on the Open ASR leaderboard.
• Gemma 4 31B Instruct (Google): 262K context, billed as a step-change improvement over Gemma 3 for small multimodal open models.
• Qwen 3.6-Max-Preview (Alibaba): Frontier preview reinforcing China's pace at the top of the leaderboard.
• DeepSeek V4 (DeepSeek): Open-source agentic model targeting near-frontier reasoning at lower inference cost.
• Trinity Large Preview (Arcee AI): 131K context window.

The hardest problem in humanoid robotics was never the AI. It was the factory.

Figure said this week that its BotQ facility now builds one Figure 03 per hour — a 24x throughput improvement in under 120 days, with more than 350 units delivered to date. That's a company-published number, not an independent audit, but the specificity makes it easy to verify or falsify, which is more than most scaling claims offer.

The interesting part is buried below the headline. Figure argues that fleet size and model quality are now coupled: more robots in the field means more real-world data feeding its Helix model, which means better robots, which means more deployable robots. The company already has commercial deals with BMW and Brookfield, and Humanoids Daily reports two more major customer announcements teased within the next 60 days.

If this works, the unit economics of physical AI flip — the bottleneck shifts from hardware to software, integration, and deployment ops, which is the regime every other humanoid maker should fear. If it doesn't, expect to see the production rate quietly stop being mentioned in future updates and reliability metrics replace throughput as the headline number. The signal to watch: a paying commercial customer beyond the existing pilots. That's when "fleet building" becomes "revenue."

Meta had a great quarter. Revenue was up 33% year-over-year for the quarter, and profits were up 61% year-over-year for the quarter. The stock fell 6% in afterhours trading on April 29, 2026.

The company raised its 2026 capex guidance to $125–145 billion, up from $115–135 billion, per Fortune. To put that in perspective: Mark Zuckerberg now plans to spend more on infrastructure in 2026 than Meta did in all of 2024 and 2025 combined. Bloomberg framed it bluntly — Zuckerberg reignited fears that the historic levels of investment he's making to catch up in the AI race won't pay off.

The contrast with Alphabet matters. Google Cloud grew 63% year-over-year to $20 billion, with a backlog that nearly doubled to $462 billion. Fortune noted that only Google convinced investors its capex was paying off. Google's stock held. Meta's didn't.

If Meta's AI ad metrics keep accelerating — impressions were up 19% year-over-year for the quarter and price per ad was up 12% year-over-year for the quarter — the $145B becomes prescient. If they don't, this becomes the moment the market started rationing AI ambition, and the second-order effect is that cheaper open models from DeepSeek, Qwen, and Kimi stop being curiosities and start being tactical weapons for everyone who can't sustain hyperscaler-level spending.

The most revealing thing about JAL's humanoid trial isn't the robots. It's the labor math that made it inevitable.

Per JAL's press release, JAL Ground Service and GMO AI & Robotics Trading will launch Japan's first humanoid airport demonstration in May 2026. Two robots from China's Unitree Robotics will move containers and operate securing levers on the tarmac at Haneda, which serves over 60 million passengers a year, Interesting Engineering reports. JAL employs about 4,000 ground handlers, per The Japan Times, and the company says the humanoid form factor was the point — unlike fixed automation, these robots fit existing infrastructure without rebuilding the airport.

The geopolitical subtext is loud and unspoken: a Japanese flag carrier is deploying Chinese robotics hardware in safety-critical infrastructure during a period when most of the developed world is trying to disentangle from Chinese tech supply chains. The trial runs through 2028. If humanoids survive a real shift on a live apron — jet blast, fuel vapor, weather, human coworkers — every airport with a labor shortage starts watching, and that's most of them. If they don't, expect a quiet pivot back to fixed automation and a much louder conversation about robot supply-chain provenance.

According to Anthropic's reporting, an internal frontier model — referred to in briefings as "Claude Mythos" — found thousands of high-severity vulnerabilities across every major operating system and browser, including a 27-year-old OpenBSD flaw discovered in a successful run that cost roughly $50 in compute. Anthropic says the model can read codebases, form exploit hypotheses, run debuggers, and chain vulnerabilities into working exploits autonomously.

These are Anthropic's own claims, withheld from broader release and gated through a defender-only program called Project Glasswing, with founding partners including AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, and Palo Alto Networks. Independent verification is limited — Heidi Khlaaf of the AI Now Institute warned against accepting the results at face value without false-positive data. But the industry partners aren't dismissing it.

If the cost of finding deep vulnerabilities really has collapsed from "expert team-years" to "fifty bucks of GPU time," the quiet assumption that critical bugs stay rare because they're hard to find just broke. The signal to watch: how many of the disclosed bugs get patched in the next 90 days. Anthropic says fewer than 1% were patched as of April 30, 2026. That number needs to move fast.

NVIDIA released Nemotron 3 Nano Omni — a 30B multimodal MoE with 256K context built for agentic workloads spanning text, image, video, audio, and documents — and the distribution story may matter more than the model. Per Latent Space's roundup, it landed same-day on OpenRouter, LM Studio, Ollama, Unsloth, Fireworks, DeepInfra, Together, and Baseten. Several hosts cited roughly 9× throughput versus comparable open omni models.

Pair this with Poolside's Apache-2 release of Laguna XS.2 (33B total / 3B active MoE, runnable on a single GPU) and you can see the shape of the next quarter: open multimodal agents with long context and serious throughput, available everywhere, at prices that make Meta's $145B look like a strategic question rather than a foregone conclusion.

If routing platforms keep pushing traffic to these models over the next few weeks, frontier labs lose pricing power on inference faster than they expected. If usage stays concentrated on closed APIs, the open ecosystem becomes a developer story rather than an enterprise one.

The biggest underappreciated story in AI right now is that China's labs are no longer just catching up. DeepSeek V4, Qwen 3.6-Max-Preview, and Kimi K2.6 all shipped in the same window, and 36Kr's analysis frames it as a sustained campaign rather than a burst — DeepSeek "no longer walks alone in the dark alley."

What changed: these labs aren't competing on raw benchmarks first. They're attacking distribution, pricing, openness, and long context — the parts of the stack enterprise buyers actually feel. Kimi K2.6 is now reportedly #1 on OpenRouter's weekly leaderboard, which means developers are choosing it in production, not just bookmarking it.

If your "default model shortlist" doesn't include at least two Chinese labs by midyear, your shortlist is out of date. The signal to watch: whether US developer platforms keep routing freely to these models, or whether political pressure starts drawing harder lines at the routing layer.

This one hasn't hit the main AI wire yet. Per Asanify's regulatory digest, the European Commission, Council, and Parliament walked out of a 12-hour Strasbourg trilogue negotiation on April 28, 2026, with no Omnibus deal — blocked on conformity-assessment rules for Annex I products like medical devices and machinery. The next trilogue is scheduled for around May 13, 2026.

The practical consequence: the August 2, 2026 high-risk deadline is unchanged. AI used in recruitment, screening, performance management, and termination sits inside Annex III. If you ship in the EU or hire EU candidates, your ATS, your interview-scoring stack, and any agentic résumé parser are in scope. Penalties for prohibited practices reach €35M or 7% of worldwide turnover.

This is one digest citing TNW and Modulos — worth independent confirmation — but the underlying regulatory fact is not in dispute. Every company that assumed the Omnibus would soften the timeline just lost that assumption. Watch the May 13 trilogue.

One humanoid an hour rolling off a California assembly line, a Unitree humanoid wrestling baggage carts on a Tokyo runway, and an Anthropic model finding a 27-year-old OpenBSD bug for the price of a sandwich. The market's verdict on Meta's $145 billion was a 6% haircut in afterhours trading on April 29, 2026 — apparently "we're spending historically" no longer counts as a strategy when the open-source Chinese lab down the street is doing it for less and the EU's regulators just walked out of the room. Stay sharp.

Forward this to the friend who's been telling you humanoids were ten years away.