The center of gravity in AI shifted east this week as China convened the World AI Conference in Shanghai and used it to launch a new intergovernmental governance body — a bid to shape the rules of AI before Washington does. Underneath the diplomacy, the money kept moving: an inference startup raised one of the year's largest rounds, robotics deals piled up, and Apple briefly wrested the "world's most valuable company" crown back from Nvidia. There was also a rough run of security news. Here's what matters for operators.
China moves to write the world's AI rules
The biggest story of the week is geopolitical, not technical. On July 16, China formally established the World Artificial Intelligence Cooperation Organisation (WAICO), an intergovernmental body headquartered in Shanghai with 29 founding member nations — most from the Global South, including Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia, and Pakistan. President Xi Jinping unveiled it at the World AI Conference the next day, his first appearance at the event since it launched in 2018, framing AI as something that "should not be a solo performance by a single country, but a symphony" of cooperation. He pointedly warned against "the misuse of national security concepts" in AI — a clear jab at U.S. semiconductor export controls — while pledging to help lower-income nations build their own AI capacity.
The subtext is a contest over standards. As the U.S. tightens chip exports, Beijing is positioning itself as the convener of inclusive AI governance, backed by middle-power and developing nations whose combined UN voting weight amplifies China's influence. No U.S. participation was mentioned. For business operators, the practical signal is fragmentation: the world is drifting toward competing AI blocs with different rules on data, safety, and model access. That directly affects which tools you can rely on — especially the cheap, capable open-weight Chinese models (like Moonshot's Kimi K3) that this ecosystem is designed to promote.
The inference economy lands a megaround
The clearest sign of where enterprise AI spending is headed came from Fireworks, which raised a $1.505 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures, and TCV, with Nvidia, Lightspeed, Bessemer, and Menlo Ventures also participating. The company says it has crossed $1 billion in annualized revenue and now serves more than 40 trillion tokens a day — and that over 95% of those tokens come from models specialized on customers' own proprietary data, not general-purpose flagships. Customers cited include the coding tool Cursor and legal-AI firm Harvey.
Fireworks' pitch — "every company must own its intelligence" — captures the shift underway: businesses are moving from *renting* a general model via API to *building* smaller, cheaper models fine-tuned on their own data and running them at scale. The operator takeaway: the durable enterprise value in AI is increasingly in the inference-and-customization layer, not the frontier model itself. For an SMB, that reinforces a build-on-your-data strategy over chasing whichever model tops this week's leaderboard.
Robotics money keeps flowing
Physical AI drew fresh capital too. Walden Robotics, a humanoid startup spun out of the Toyota Research Institute in January, emerged from stealth with roughly $300 million in funding at a $1.1 billion valuation, co-led by Deviation Capital and Toyota, with Nvidia, Boeing, Samsung Ventures, and CoreWeave Ventures among the backers. Co-founder Russ Tedrake says Walden is already selling humanoids to customers, including a pilot at a North American Toyota plant where a robot works eight-hour shifts loading and unloading car parts, cleaning machinery, and kitting for assembly. In Germany, Munich-based Microagi raised €55 million — the largest seed round in German history — for a business that collects data from live production lines and then leases automation hardware (often from Chinese makers Unitree or UBTech) into factories. The through-line: humanoids are moving from demo reels to paid shift work on real production floors, and manufacturing SMBs are the near-term customers.
Big Tech jockeys for position
Apple briefly reclaimed the title of world's most valuable company on July 17, edging past Nvidia intraday before Nvidia closed slightly ahead again — both hovering near a $4.9 trillion market cap, neither yet across the $5 trillion line. The move reflects a rotation as investors reward Apple (up ~23% in 2026) over Nvidia (up ~7%) amid fresh doubts about AI-capex payback. Meanwhile Meta hired senior AWS compute executive Dave Brown and is reportedly weighing an entry into cloud services to help absorb a planned $125–145 billion in capital spending this year. And Microsoft is preparing "Project Perception," a cybersecurity tool that blends models from Microsoft, OpenAI, and Anthropic into a single multi-model defense product, expected this month.
A hard week for security and IP
Several incidents landed at once. Music-AI startup Suno disclosed a breach exposing source code and details of its training data — reportedly including material sourced from YouTube Music, Deezer, Genius, and podcast libraries — which could deepen the copyright scrutiny already dogging generative-audio firms. Coca-Cola subsidiary Fairlife halted U.S. production after a cyberattack compromised manufacturing systems, a reminder that AI-era operational tech remains a soft target. And Apple sent cease-and-desist letters to roughly 40 former employees now at OpenAI, accusing the lab of soliciting hardware trade secrets — an escalation in the talent war between the two.
Quick Takes
Google's Gemini 3.5 Pro slips again. Alphabet has pushed the broader release out by months after the model underperformed on coding and long-horizon reasoning; Alphabet shares fell about 4% on the news. It remains in limited enterprise preview.
Kimi K3's open weights are coming July 27. Moonshot's 2.8-trillion-parameter model — already ranked #1 on the Frontend Code Arena, ahead of Claude Fable 5, though still behind Fable 5 and GPT-5.6 Sol on overall benchmarks — will publish full open weights, which could pull self-hosted coding costs down sharply.
OpenAI's GPT-Live is now the default voice. The full-duplex model that listens and speaks at once (delegating hard queries to a frontier model in the background) has rolled out globally in ChatGPT, replacing Advanced Voice Mode.
Germany tightens the screws on AI search. Germany's media regulator ruled that Google's AI Overviews and Perplexity are subject to German media law — an early sign that AI-generated answers will face publisher-style regulation in Europe.
NVIDIA courts Japanese industry. Japanese robotics and manufacturing leaders are building on NVIDIA's Cosmos platform and Nemotron open models to develop industry-specific "physical AI," reinforcing the enterprise pull toward domain-tuned systems.
What This Means for Your Business
The week's dominant theme for operators is specialization over supremacy. Fireworks' billion-dollar run rate and Xi's promotion of open models point the same direction: the winning move for most companies is no longer picking the single "best" frontier model, but taking a capable, affordable base — increasingly an open-weight one — and tuning it on your own data for a specific job. If you haven't yet, identify one or two high-volume, well-defined tasks (support triage, document extraction, code review) where a fine-tuned small model could replace expensive general-purpose API calls, and pilot it.
Second, plan for a fragmenting model landscape. With China building a governance bloc around open models and Europe extending media law to AI answers, the regulatory and availability picture will differ by region. Don't architect your product around a single vendor or a single jurisdiction's assumptions; keep an abstraction layer so you can swap models — Western or Chinese, closed or open — as price, performance, and compliance shift. The Kimi K3 open-weights drop on July 27 is a concrete date to evaluate: it may cut your coding-assistant costs materially, but running a Chinese model may carry data-governance questions worth vetting first.
Third, treat security as a first-class AI cost. The Suno breach and the Fairlife production halt are reminders that as you wire AI into more systems, you expand your attack surface. Microsoft's multi-model "Project Perception" hints at where defense is going, but the basics matter more: inventory where AI tools touch sensitive data and operational systems, and assume any training data or source code you feed a third-party model could someday leak. Finally, if manufacturing or logistics is your world, the Walden and Microagi deals signal that leased, general-purpose humanoids are becoming a real line item — worth a pilot conversation this year, not a "someday" one.