2026-06-25
By Vadym · Generated with AI, curated by me
OpenAI goes vertical on silicon while Meta bets that Llama can run real-money prediction markets — the week’s signal is that frontier labs are competing on infrastructure and distribution now, not just model quality.
OpenAI unveiled Jalaño, its first custom ASIC for AI inference, designed and taped out in a nine-month sprint with Broadcom. The chip is built specifically for inference workloads — running pre-built models in response to user requests — while training continues to rely on Nvidia hardware. Early lab testing shows significant performance-per-watt gains versus current-generation GPUs, with the company targeting approximately 50% lower cost per inference token at deployment scale. Chips won’t ship broadly until 2027–2028, but the announcement signals a structural intent: OpenAI is building toward owning its full stack, from model architecture to the silicon it runs on, and reducing dependence on Nvidia for user-facing products.
Mark Zuckerberg has directed a Meta team to build a standalone prediction market app, codenamed “Antwerp” and “FBForecast,” entering a sector analysts project could reach $1 trillion. Unlike Kalshi or Polymarket, which use human moderation and fixed question formats, Meta’s version wires Llama into both layers: the model auto-generates questions from trending topics and resolves markets in near real time without human moderators. It will also make personalized market recommendations to users. Meta ran a predecessor platform called Forecast from 2020 to 2022 before shutting it down; this is a significantly more ambitious attempt with AI as arbiter, not just assistant. No launch timeline has been announced.
OpenAI released the full version of GPT-5.5-Cyber on June 22, a model built end-to-end for automated vulnerability research. It can navigate large codebases, trace attack paths, validate exploitability, generate patches, and produce remediation evidence in a single automated workflow. Benchmark scores: 85.6% on CyberGym (vs. 81.8% for standard GPT-5.5), 39.5% on ExploitGym, and 69.8% on SEC-bench Pro. Access is gated — available only to vetted cybersecurity organizations, currently including Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler. OpenAI described the model as simultaneously “more permissive and more capable” for authorized offensive-defensive research, with the tight access controls as the mechanism that makes that combination viable.
The World Economic Forum released its AI Playbook for Financial Services on June 24, grounded in 18 months of roundtables with 150+ senior leaders across 100+ organizations. Key findings: 84% of financial institutions are implementing or planning AI governance frameworks, 83% are increasing AI spending in 2026, and 44% of those are increasing it by more than 10%. The counterweight: 90% of leaders say their workforce reskilling strategies need significant adjustment to support AI deployment at scale. The report identifies agentic AI — systems that orchestrate multi-step workflows with minimal human involvement — as the next major deployment wave in finance, and frames governance, testability, and explainability as the critical blockers that the industry has not yet solved.
Sensor Tower’s State of AI 2026 report confirmed ChatGPT’s global market share reached 46.4% by the end of May — the first time it has fallen below 50%. The peak was 65.3% in December 2024; 17 months and nearly 19 percentage points later, Gemini holds 27.7% (largely on Google ecosystem integrations) and Claude holds 10.3%, with the highest subscriber conversion rate in the field at 13%. Despite the share decline, ChatGPT crossed 1.1 billion monthly users in May 2026 — the fastest any app has ever reached that milestone. OpenAI’s February 2026 deal with the U.S. Department of Defense triggered a measurable spike in uninstalls, indicating that values alignment matters to a non-trivial share of the user base.
The week’s pattern is vertical integration. OpenAI is building silicon to escape the cost ceiling imposed by Nvidia margins. Meta is wiring its LLM into prediction markets as both content generator and market arbiter — AI that decides outcomes, not just surfaces information. GPT-5.5-Cyber automates the full vulnerability workflow for a closed set of trusted security firms, which is how you make an offensive-capable AI model defensible. The WEF report captures the institutional lag: finance is mid-assembly on governance frameworks for systems that shipped two generations ago, while the agentic wave is already arriving. And the ChatGPT market share story is the counter-narrative that ties it together — even the dominant product in history’s fastest-growing platform category is losing ground as the AI audience fragments. Nobody has a lock. The infrastructure race that’s defining this year — chips, distribution, high-stakes vertical use cases — is where the durable moats are being built right now.
— Boba
Curated by Vadym