AI News Roundup — June 24, 2026
OpenAI unveiled 'Jalapeño' — its first custom AI inference chip built in collaboration with Broadcom — joining Google TPUs, Amazon Trainium, and Microsoft Maia in the custom silicon race to cut Nvidia dependence. Alphabet stock fell 7.2% intraday as two landmark researcher departures in a single week (Noam Shazeer to OpenAI, John Jumper to Anthropic) hit sentiment hard, while Gemini 3.5 Pro is now down to 6 days to meet Sundar Pichai's June commitment.
OpenAI Unveils 'Jalapeño' — Its First Custom AI Inference Chip Built With Broadcom
OpenAI unveiled its first custom-built AI processor, named Jalapeño, designed and manufactured in collaboration with Broadcom. The chip was designed specifically for inference — running pre-trained AI models in response to user requests — with early results showing significantly better performance-per-watt than current state-of-the-art alternatives. The chip was built in a sprint development cycle, with OpenAI's own AI models assisting in the design process. OpenAI president Greg Brockman: "We have a deep understanding of the workload — we've been looking for specific workloads that are underserved and asking how we can build something that accelerates what's possible." Jalapeño is specifically optimized for real-time coding model inference (a clear Codex reference), though pre-training will still rely on Nvidia hardware. The chip joins Google's TPUs, Amazon's Trainium, Microsoft's Maia 200, and Apple's neural engines in the frontier-lab custom silicon race. OpenAI is explicitly framing the move as a full-stack strategy: "OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience." The strategic implication for Nvidia: OpenAI's inference workloads — including the 900M+ weekly ChatGPT users and Codex deployments — represent a massive sustained GPU demand that Jalapeño may progressively displace.
TechCrunch · OpenAI · Reuters · Bloomberg
Alphabet Stock Falls 7.2% — Two Google AI Giants Leave in One Week
Alphabet stock fell as much as 7.2% intraday on June 24 — its steepest single-day decline since February 2026 — following the combined weight of two landmark researcher departures in a single week: Noam Shazeer (Transformer co-author, Gemini VP of Engineering) to OpenAI on June 18, and John Jumper (Nobel Prize co-laureate, AlphaFold lead) to Anthropic on June 20. The market is not simply reacting to two people leaving — it is pricing in the downstream implications for Gemini 3.5 Pro (still unshipped with 6 days left in Google's June commitment), the architectural knowledge of Gemini's weaknesses that now resides at both rival labs, and a talent signal that raises questions about Google's culture, compensation, and competitive position in the frontier AI race. Google's $4.46 trillion market cap and dominant Search advertising revenue give it structural resilience, but the week demonstrated that even the world's most valuable cloud company cannot treat frontier AI talent as retained by inertia when OpenAI and Anthropic are competing aggressively for exactly those researchers.
Reflection AI — The '$25 Billion DeepSeek of the West' Explained
The company behind SpaceX's $6.3 billion Colossus 2 compute deal is Reflection AI, founded by Misha Laskin and Ioannis Antonoglou — both former Google DeepMind veterans (Antonoglou helped build AlphaGo). The company is backed by Nvidia, Sequoia, and Lightspeed, and is raising at a $25 billion valuation. Its thesis: American governments, banks, defense contractors, and large enterprises want frontier-capable AI with open weights — but will not use closed US labs for sovereignty reasons and will not use Chinese open-weight models for security reasons. Reflection is positioning as the third option: American, open-weight, and frontier-scale. The company has ties to the Department of Energy's Genesis Mission and Pentagon AI programs. Its code research agent Asimov has been on a waitlist since early 2026. No public frontier model has shipped yet. The $6.3 billion Colossus 2 compute commitment is the most concrete signal that a frontier training run is imminent. Tom's Hardware notes that Nvidia is both the chip supplier and a major investor in Reflection AI — creating a circular dynamic where Nvidia sells the picks and owns part of the mine.
Build Fast With AI · Tom's Hardware
Gemini 3.5 Pro — 6 Days Left, Leaks Suggest Reasoning Gap vs Fable 5 and GPT-5.6
Gemini 3.5 Pro has 6 days remaining in Sundar Pichai's June commitment window and remains in limited Vertex AI enterprise preview only. Polymarket currently prices a 50–55% chance of a June 30 general availability. Early leak reports from developer previews suggest Gemini 3.5 Pro shows a reasoning gap versus both Fable 5 and the pending GPT-5.6 on complex multi-step tasks, though it performs well on multimodal benchmarks. The window matters more than usual: Gemini 3.5 Pro is the only major product that could give Google a positive AI narrative before Q3, after losing two flagship researchers in a single week. Missing the June deadline — or shipping to lukewarm benchmarks — would compound an already difficult narrative cycle.
Fable 5 Return Signals Emerge in Claude Code and AWS — Unverified But Notable
RuntimeWire and explainx.ai reported that some Claude Code version 2.1.190 users are seeing Fable 5 appearing as an available model option again, and that AWS Bedrock documentation has been updated to include claude-fable-5 as a listed model. Neither Anthropic nor the Commerce Department has made an official announcement confirming restoration. Explainx.ai calls the reports "unverified — treat as rumor until official announcement." The signals are nonetheless significant: they suggest that a technical compliance path is ready to deploy, most likely involving data retention requirements (Anthropic's AWS blog post notes users must opt into data sharing via the Data Retention API before accessing Fable 5). If the signals are genuine, an official announcement may be imminent. Polymarket's July 1 restoration contract is now trading at approximately 70% probability.
Meta Leads $850 Billion Data Center Lease Boom With Microsoft — AI Infrastructure Race Accelerates
Meta and Microsoft are leading a global surge in data center leasing, with $850 billion in combined signed commitments reported. The figure is part of a broader AI infrastructure boom that has now absorbed over $400 billion in debt and equity financing in roughly 90 days, and is now manifesting as concrete signed lease agreements for power, land, and rack space. The data point underscores the same dynamic driving FERC's June 18 grid orders and Nvidia's record $25 billion bond: the physical infrastructure requirements for next-generation AI are outpacing every prior technology buildout, and the companies leading the AI race are simultaneously the world's largest capital allocators.
Bloomberg Tech / Open Interest 6/24/2026
OpenAI Acquires Ona — Persistent Sandbox Environments for Codex
OpenAI acquired Ona, a startup specializing in persistent compute sandbox environments, to integrate directly into Codex. The acquisition addresses one of Codex's key limitations for complex long-running tasks: agents lose execution context when a session ends. Persistent sandbox environments mean Codex agents can maintain stateful compute sessions across multiple working sessions — resuming where they left off, retaining installed dependencies, and tracking changes over time. This is a significant architecture improvement for enterprise Codex workflows involving large codebases, multi-day projects, and automated CI/CD pipeline integration. The acquisition underscores OpenAI's vision for Codex as a full-stack software development environment rather than an autocomplete tool.