Tech Digest – July 29, 2026
The Open-Source Frontier
Moonshot AI Opens Kimi K3 — First 3-Trillion-Class Model Available for Download
Moonshot AI released the full weights of Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with vision capabilities and a million-token context window. It is the first open model at the 3-trillion-parameter scale, selecting 16 of 896 experts per token. Within days of release, K3 logged 99,000 downloads on Hugging Face and drove a sixfold jump in Moonshot’s sales, with the company now chasing a $50 billion valuation.
The release carries a geopolitical edge: the White House alleged the model was trained on export-controlled Nvidia silicon, and Moonshot’s next model, Kimi K4, reportedly requires more Blackwell chips than Beijing can currently source. Amazon, reading the competitive landscape, retired most of its Nova models to redirect scarce compute to roboticist Pieter Abbeel’s new lab. Meanwhile, enterprise buyers are increasingly going à la carte across cheaper AI models — some Chinese — shifting who holds market power.
Note: A year ago, frontier intelligence required a subscription to San Francisco. Now it’s a download. For any institution weighing data sovereignty against capability, the trade-off just collapsed.
Sources: Bloomberg, Hugging Face, Business Insider, WSJ
AI Governance & the August Deadline
Washington Scrambles Before August 1 — OpenAI and Anthropic Ally, 1,100 Staffers Push Back
The Trump administration is nearing a voluntary framework for pre-release review of frontier AI models, with open-source treatment still unresolved ahead of an August 1 deadline. Rivals OpenAI and Anthropic quietly allied to shape the outcome, while over 1,100 employees at both companies petitioned the government to “deliberately pace the frontier” — days after OpenAI’s own models exploited vulnerabilities on Hugging Face. The breach accelerated Nvidia’s launch of an Open Secure AI Alliance, which argued that Hugging Face contained the damage precisely because it ran open weights on its own infrastructure.
The CEOs staked out positions publicly. Altman urged Washington to restrict Chinese access to American chips. Amodei clarified that Anthropic never sought to ban open weights — only to withhold chips from adversaries and mandate safety testing. Zuckerberg argued that the real question is not whether superintelligence arrives, but who gets one.
Note: The August deadline is US-specific, but the framework it produces will set the baseline EU regulators benchmark against. If you’re planning AI procurement on a 2027 timeline, the regulatory ground is shifting under you this week.
Sources: The Information, Bloomberg, CNBC, Anthropic, Nvidia, WSJ
Semiconductor Sovereignty
Shanghai Mass-Produces Domestic Lithography, CXMT Debuts at +466% — China’s Chip Supply Chain Arrives
Shanghai began mass-producing domestic immersion DUV lithography machines, giving China its first industrial-scale alternative to ASML’s equipment. Memory chipmaker CXMT debuted on the Shanghai exchange with a 466% first-day surge, closing at roughly $488 billion in market value — making it China’s most valuable listed company after a $9.8 billion IPO, Asia’s largest this year.
The incumbents aren’t fading — SK Hynix reported 257% revenue growth on HBM demand — but European and Korean chipmakers sold off sharply anyway. ASML dropped 3.4% and the Kospi fell 11%, with analysts citing “AI fatigue.”
Note: Markets read this as AI fatigue. What it actually looks like: cheaper memory, redundant lithography, and a second global chip ecosystem emerging faster than expected. De-risking through diversification — the vocabulary the EU has been using about supply chains — is now happening whether Brussels planned for it or not.
Sources: Tom’s Hardware, CNBC, CNBC, Bloomberg
Infrastructure at Scale
Over $275 Billion in New AI Infrastructure Commitments — and Microsoft Is Still Rationing
Nvidia is in talks to backstop $250 billion in financing for OpenAI’s 10-gigawatt Ohio campus, a decommissioned uranium-enrichment site roughly 50 miles south of Columbus. Meta and BlackRock committed $14 billion to a gigawatt facility in El Paso. AMD locked up 500 megawatts at Core Scientific. On the regulatory side, the EPA ruled that power plants serving only data centres can bypass the Acid Rain Program, removing a barrier to rapid construction.
The workforce bottleneck is getting its own investment: Google, Meta, and BlackRock are funding apprenticeship programmes for thousands of electricians, whose remote data-centre work now pays 42% more than metro equivalents. Yet even $190 billion in cumulative capital expenditure hasn’t been enough for Microsoft, which is down 24% this year — the company is rationing Azure capacity, prioritising frontier AI labs over commercial customers and renting infrastructure from competitors to fill the gap.
Note: When the company that spent $190 billion on infrastructure is renting capacity from rivals, the bottleneck isn’t capital — it’s physics. Cloud migration timelines built around Azure availability should be stress-tested against the reality of rationed allocation.
Sources: WSJ, Meta, The Block, Reuters, NYT, Business Insider
Satellite Images Confirm Iranian Strikes on Amazon’s Bahrain Data Centres
Satellite imagery from the European Space Agency’s Sentinel-2 constellation confirmed damage to two Amazon Web Services data centres in Bahrain — in the towns of Zallaq and Askar — following Iranian missile strikes. The IRGC claimed the attacks were retaliation for US strikes on an Iranian nuclear facility. The IRGC also released high-resolution imagery of the damage. Amazon has not commented.
Note: When compute becomes terrain worth striking, disaster recovery plans need a new section. Geographic distribution of cloud workloads is no longer just a latency decision.
Physical AI & Trade Barriers
FCC Bars Chinese Humanoid Imports — Baidu Begins Testing Robotaxis in London
The FCC immediately banned new imports of Chinese-made humanoid robots and quadrupeds, citing infrastructure security risks. China holds an estimated 85% of the global humanoid market; the ban directly impacts Unitree, which recently integrated Nvidia’s Blackwell chip as a robot brain. On the other side of the trade barrier, Baidu began testing its Apollo Go robotaxis in London alongside Waymo and Wayve, bringing Chinese autonomous vehicle technology onto European-adjacent roads.
In Cape Coral, Florida, a city reversed plans to mount AI code-enforcement cameras on garbage trucks — designed to autonomously grade residents’ lawns — after public backlash. An early data point for any municipality weighing automated compliance enforcement.
Note: First time a major economy has banned an entire category of physical AI products. Brussels is watching. The EU’s own robotics pipeline has Chinese dependencies it hasn’t audited yet.
Sources: Reuters, Washington Post, TechCrunch
Workforce & Institutional Risk
AI Erases Call-Centre Jobs at Microsoft, Uber, and CBA — Half the Profession Exposed by 2030
AI-driven automation is eliminating customer service roles at scale. Microsoft’s support workforce dropped from roughly 50,000 to 40,000 — with AI saving an estimated $750 million annually. Uber cut 10% of its community operations team. Commonwealth Bank of Australia eliminated 120 roles despite earlier union pushback. A Forrester analyst estimated that nearly half of all customer service positions will be affected by 2030.
Separately, Visa cut 2,600 jobs — 7% of its workforce — to reinvest in stablecoin and digital payment infrastructure, signalling that the automation pivot extends beyond AI into broader fintech restructuring.
Note: Many public institutions outsource first-line citizen contact to the same call-centre providers losing these contracts. The workforce shift doesn’t stop at the private sector — it arrives through the supply chain.
40 Lawsuits Now Allege Chatbots Contributed to User Deaths
Bloomberg tracked approximately 40 lawsuits filed against chatbot makers since late 2024, alleging their products contributed to user deaths, suicides, and acts of serious violence. At least 20 deaths are cited across the filings, many involving minors. The states of Florida, Kentucky, and Utah have also filed suits. Most cases target OpenAI, Character.AI, and Google over alleged failures in safety design and content moderation.
Note: Forty lawsuits in two years, many involving minors. Case law is forming around a question that matters to every public-sector chatbot deployment: who is liable when the model gives dangerous advice?
Sources: Bloomberg
Today’s digest is a fracture map. The frontier is splitting — open and closed models, US and Chinese chips, American and Chinese robots — and each split doubles the decisions an institution has to make. The common thread isn’t acceleration anymore. It’s proliferation: of capability, of supply chains, of risks, and of the questions that land on a decision-maker’s desk.
