Good morning. It's Thursday, May 28, and we're covering a $100 million bet on modular fusion, the 40% collapse hiding under Asia's AI export boom, and the botnet that poisoned 300+ GitHub repos.

Yesterday's Real vs AI Results + the "Best Model" Poll ↓

Hackers quietly poisoned more than 300 GitHub repositories by targeting developers instead of the software itself. CrowdStrike, working with Google and Shadowserver, dismantled parts of the Glassworm botnet used in open-source supply chain attacks. The group relied on Solana, BitTorrent, Google Calendar, and VPS infrastructure to control infected machines over a two-year campaign. The takedown follows recent attacks tied to compromised open-source tools, including Axios and the Mini Shai-Hulud campaign.

The OpenAI Foundation is directing $250 million toward cushioning workers from AI-driven disruption. OpenAI said the money will support labor-market research, worker protections, and experiments for distributing AI-generated wealth more broadly. The initiative includes studies in low- and middle-income countries, support for unemployment and wage-loss systems, and policy models like sovereign wealth funds. Reuters reported the foundation now holds a 26% stake in OpenAI’s commercial business following last year’s restructuring.

A new benchmark from Datacurve claims leading AI coding models are far less equal than existing leaderboards suggest. DeepSWE ranks OpenAI’s GPT-5.5 at 70%, far ahead of rivals, while also alleging Scale AI’s SWE-Bench Pro misgraded nearly one-third of reviewed trials. The study found Claude Opus exploited Git history inside benchmark containers to retrieve answer commits, inflating some scores by up to 25%. Datacurve said DeepSWE spans 113 tasks across 91 repositories and requires roughly 5.5 times more code than SWE-Bench Pro.

While governments push to diversify chip supply chains, NVIDIA says it will spend up to $150 billion annually in Taiwan, calling the island the “epicentre” of the AI revolution. CEO Jensen Huang unveiled a new Taipei headquarters expected to open by 2030 and employ 4,000 people. The investment strengthens ties with TSMC, Foxconn, Wistron, and Quanta as demand for AI chips and servers accelerates. Huang said NVIDIA’s Taiwan spending has jumped roughly tenfold in five years, from about $10–15 billion annually.

ElevenAgents by ElevenLabs enables you to easily build intelligent, real-time conversational AI agents. By using your own knowledge base, SOPs and policies, ElevenAgents creates agents that sound like your brand in over 70 different languages at low latency and in minutes.

Ship voice agents for sales, support, and operations. Configure, deploy, and monitor natural, human-sounding agents that scale your team without scaling headcount.

Try ElevenAgents from ElevenLabs here.

New DeepSWE coding benchmark crowns GPT-5.5 over Opus 4.7, highlighting real-world coding performance, speed, and cost gaps.

When you need to get real work done, which model do you actually reach for first?

Sharing what everyone said tomorrow.

Princeton spinout Thea Energy has raised an oversubscribed $100 million Series B led by U.S. Innovative Technology Fund, bringing its total private funding to $130 million. The company is developing a stellarator fusion reactor that uses hundreds of smaller, software-controlled magnets instead of the large custom-shaped magnets common in other designs.

The new funding will support magnet manufacturing expansion and construction of Eos, a demonstration reactor planned for 2027, with a commercial system called Helios targeted for 2034. Thea argues its “pixel-like” magnet approach could simplify one of fusion energy’s most expensive engineering challenges, though the company still relies on larger external magnets for core plasma confinement. → Read the full article here.

Japan, South Korea, and Taiwan are posting record export growth and corporate profits, driven largely by surging demand for artificial intelligence chips, servers, and semiconductor equipment. But outside the AI supply chain, much of the region’s industrial base is weakening as Chinese manufacturers gain ground in sectors including cars, batteries, machinery, and chemicals.

Taiwan’s non-AI exports have fallen 40% since 2022, while South Korea and Japan are seeing stagnation or decline across traditional manufacturing industries. Governments are doubling down on semiconductor investment despite growing dependence on exports and a narrow group of overseas buyers, particularly the U.S. and China. → Read the full article here. (Paywall)

Alibaba’s Qwen team said its Qwen3.5 model reached a record 580 tokens per second for agentic AI workloads using the TokenSpeed inference engine on NVIDIA GPUs. The benchmark, highlighted in a new PyTorch Foundation blog post, was achieved through collaboration with LightSeek, NVIDIA AI, Mooncake, and TRI DAO, which contributed FlashAttention-4 (FA4) optimizations.

The result underscores the growing competition to improve inference efficiency for large language models, where speed and lower compute costs are becoming as important as model quality. → Read the full article here.

Cognition Hits $26B Valuation: AI coding startup Cognition raised $1 billion as annualized revenue hit $492 million and usage grew over 10x this year.

Meta Sells Ad-Free Plans: Facebook and Instagram will charge UK users £2.99 monthly to remove ads after pressure from regulators.

Robinhood Opens AI Trading: Robinhood launched AI agents that can trade stocks and spend through virtual credit cards with user-set limits.

Taiwan Probes NVIDIA Chip Smuggling: Taiwan detained three suspects accused of routing restricted Nvidia AI servers through Japan into China.

Tensormesh Lands NVIDIA, AMD Backing (Paywall): AI inference startup Tensormesh raised a $20 million seed extension to cut AI compute costs.

Image 2 was AI. Here’s how you voted:

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team