The Signal — May 1, 2026
The AI industry is building faster than ever while simultaneously deciding who gets to use what it builds. Two OpenAI stories land back-to-back that capture this tension, and Samsung's record earnings show the hardware layer straining to keep up.
OpenAI Restricts Its Cyber Tool — Days After Mocking Anthropic for Doing the Same
Last week, Sam Altman accused Anthropic of "fear-based marketing" for limiting access to Mythos, its cybersecurity-focused AI tool. Days later, OpenAI announced that GPT-5.5 Cyber will ship with the same restrictions, available only to vetted "critical cyber defenders" rather than the general public.
WIRED reports that OpenAI is rolling out a new "Advanced Account Security" requirement for anyone wanting access to the Cyber program, taking effect June 1. The feature bundles passkey-based sign-in, tighter session controls, and restricted account recovery, essentially the same kind of gated access Anthropic built for Mythos that Altman mocked on a podcast.
When your most capable models can find zero-days as easily as they can patch them, broad access becomes a genuine liability. Both companies arrived at the same conclusion: the most dangerous AI capabilities need distribution controls, regardless of what sounds good in an interview. BusinessToday notes that Altman has not publicly addressed the contradiction.
Sources: TechCrunch · WIRED · BusinessToday
OpenAI Hits 10GW Compute Target Three Years Ahead of Schedule
While one arm of OpenAI figures out who should access its tools, another is making sure those tools have enormous power behind them. The company announced it has signed contracts for 10 gigawatts of AI computing capacity in the United States, a target originally pegged to 2029 when the Stargate project was unveiled in January 2025.
The NVIDIA letter of intent from March has converted into binding contracts, and Bloomberg independently confirmed the deals. For context, 10GW is roughly the electricity consumption of a mid-sized European country. OpenAI is locking in power generation capacity on a national scale, three years ahead of its own timeline.
The acceleration reflects raw demand for training next-generation models and competitive pressure from Google, Meta, sovereign AI programs, and well-funded startups chasing the same hardware. Falling behind on compute at this scale isn't a strategic inconvenience; it threatens your ability to compete at all. Whether the power grid and supply chains can actually deliver 10GW remains the harder question.
Sources: OpenAI Blog · Bloomberg
Samsung Posts Record Quarter as AI Memory Demand Goes Parabolic
AI's physical bottleneck is in memory chips, and Samsung's Q1 2026 earnings make that obvious: operating profit surged roughly 753% year-over-year to 57.2 trillion KRW (roughly $41.6 billion), with semiconductor division profits jumping nearly 50-fold.
Revenue hit 133.9 trillion KRW (~$97 billion), driven overwhelmingly by high-bandwidth memory (HBM) and other AI-optimized chips. CNBC reports that every major hyperscaler is competing for allocation, and Samsung is warning that supply constraints will deepen through 2027.
This connects directly to the OpenAI compute story. When one company signs 10GW of compute contracts, someone has to fill those racks with silicon. Samsung, SK Hynix, and Micron control the supply, and all three know it. SamMobile notes that Samsung is accelerating fab expansion plans, but new capacity won't come online until late 2027 at the earliest. The demand curve is outrunning supply with no intersection in sight.
Sources: Reuters · CNBC · SamMobile
On the Editor's Desk
Held or killed today: A White House story opposing Anthropic's Mythos expansion failed our freshness gate and relied on unnamed administration officials. The Anthropic $50B/$900B funding round, EU AI Act delays, and Google's Pentagon AI deal all ran in yesterday's edition. We blocked an NVIDIA NemoClaw piece for overlap with recent coverage, and Anthropic's BioMysteryBench announcement got cut as a self-reported vendor benchmark with cherry-picked framing. Six stories killed, three survived. That's a normal ratio for a space generating more press releases than breakthroughs.
