Seven Models from Scratch

Capability & Integration

• Microsoft launched seven in-house models under the MAI family at Build 2026 (June 2), with the company explicitly framing them as a hedge against OpenAI dependence. All seven were trained from scratch on commercially licensed data with no distillation from any third-party model family, including MAI-Thinking-1 (35B active-parameter MoE, 256K context, 97% on AIME 25 per Microsoft’s internal evals) and MAI-Code-1-Flash (5B parameters, native to GitHub Copilot and VS Code).

• OpenAI’s ChatGPT Ads Manager opened to all U.S. businesses in May, with conversion-optimized campaigns rolling out June 5 and the UK pilot confirmed live June 6; CPC/CPM bidding, no minimum spend. It is the first time OpenAI has embedded a full advertising surface directly in its consumer product.

• NVIDIA released Nemotron 3 Ultra 550B on June 4; model release trackers note new models are now arriving roughly every two days across providers including Google, Alibaba, and Anthropic — a pace that makes comparative evaluation increasingly impractical for most practitioners.

• A circulating analysis estimates Anthropic and OpenAI may be spending more than $1,000 in compute for every $100 in consumer revenue they collect, driven by the gap between inference costs and subscription pricing; OpenAI is projecting $14 billion in net losses for 2026.

Robotics

• Boston Dynamics’ electric Atlas has begun initial deployments to Hyundai’s Robotics Metaplant Application Center and Google DeepMind — limited pilots; the company states all 2026 capacity is committed and no new customers will be onboarded until early 2027. This is a controlled first-customer rollout, not broad industrial deployment.

• Figure AI’s BotQ factory is reportedly producing Figure 03 at 1 robot per hour, a production-rate milestone, while Figure 02 units previously accumulated 1,250+ hours in a real BMW production pilot. These are production-ramp and pilot-hours figures, not deployment-at-scale figures.

• The AGIBOT WORLD CHALLENGE at ICRA 2026 in Vienna brought 526 teams from 27 countries to compete on embodied AI tasks — advancing from simulation-only to real-hardware testing rounds. This is a research benchmark competition; results do not translate directly to deployment capability.

• Amazon’s warehouse robot fleet crossed 1 million units, with its DeepFleet AI system improving travel efficiency 10% across the network — the largest deployed autonomous robot fleet in industrial operation, though these are purpose-built mobile shelving units, not general-purpose humanoids.

Hardware & Supply Chain

• TSMC’s CoWoS advanced packaging is sold out through the end of 2026, with capacity scaling from ~35,000 wafers/month in late 2024 to a projected 130,000/month by year-end — still insufficient to clear the backlog. NVIDIA has pre-booked 800,000–850,000 CoWoS wafers for 2026, representing more than half of projected total output; other customers, including Google, have been turned away from additional allocations.

• At COMPUTEX 2026 (June 1–5), NVIDIA announced its Vera CPU has entered full production — 88 Olympus cores, 1.2 TB/s memory bandwidth — targeting reinforcement learning and agentic inference workloads in the Vera Rubin data center platform.

• NVIDIA also announced RTX Spark, a hybrid Grace CPU + Blackwell GPU superchip developed with MediaTek for consumer Windows PCs, shipping later this year from Dell, HP, ASUS, Lenovo, and MSI — extending NVIDIA’s silicon presence from hyperscale clusters to the consumer edge.

• TSMC will integrate NVIDIA AI into its fab operations to improve yield, energy efficiency, and turnaround time — a closed loop in which NVIDIA chips are used to optimize the production of NVIDIA chips.

Environmental & Cultural Impact

• A UN University report published June 3 found that in 2025, global data centers consumed 448 TWh of electricity — more than all but 10 countries — producing approximately 208 million tons of CO₂, roughly equivalent to Argentina’s annual emissions. Generating that electricity required 1.2 trillion gallons of water.

• The report projects that by 2030, AI data center electricity use will reach 945 TWh (nearly three times the combined annual use of Pakistan, Bangladesh, and Nigeria combined), with a water footprint equivalent to the basic domestic water needs of all 1.3 billion people in Sub-Saharan Africa and a land footprint exceeding 14,500 km² — roughly twice the Jakarta metropolitan area.

• Individual large data centers can consume up to 5 million gallons of water per day; a typical AI data center draws as much power as 100,000 households according to IEA estimates.

• Lawmakers in more than 30 U.S. states have introduced over 300 bills in 2026 on data center policy — moratoriums, tax incentives, grid priority rules — representing a fragmented state-level response in the absence of a federal framework.

AI in the Wild

Spencer Pratt’s AI-generated campaign ads for the Los Angeles mayoral race went viral around the June 2 California primary: one video depicts the former Hills cast member as Batman fighting a Joker-painted Karen Bass in dystopian Los Angeles, accumulating over 5 million views on X. Origin is confirmed — Pratt reposted supporter-made AI ads throughout the campaign. That a first-time candidate with no political background placed competitively in a major-city primary while running partly on AI attack content suggests the format has crossed from novelty to tactic.

Takeaway