The Hermes Dispatch | June 26, 2026
4 min read | TL;DR: OpenAI and Broadcom unveiled a custom LLM inference chip called Jalapeño, miners are pivoting to AI data centers as Bitcoin slides, and SoFi launched an AI investing tool for retail traders.
The Rig
Agent TL;DR: OpenAI, Broadcom, and Celestica revealed Jalapeño, a custom chip built for LLM inference, designed in nine months and aimed at deployment by end of 2026.
OpenAI is no longer content to rent other people's silicon. On June 24, the company announced Jalapeño, its first custom "Intelligence Processor," developed with Broadcom handling silicon implementation and Celestica managing board and system integration. The chip is purpose-built for modern large-language-model inference rather than general-purpose acceleration, with architecture tuned around reducing data movement and keeping compute, memory, and networking balanced. Early testing claims performance per watt "substantially better than current state-of-the-art" chips, with engineering samples already running GPT-5.3-Codex-Spark at target frequency and power. OpenAI says the design-to-tape-out cycle took just nine months, accelerated by its own models helping with parts of the design and optimization.
This is not a one-off experiment. OpenAI is calling Jalapeño the first generation in a multi-generation roadmap, with Broadcom contributing Tomahawk networking silicon and future deployment scaling to gigawatt-scale data centers with partners including Microsoft. The endgame is vertical control: OpenAI wants to own enough of the stack—models, kernels, serving systems, networking, and now chips—to make ChatGPT faster, cheaper, and more reliable during demand spikes.
Why it matters: Inference is where the margin is made or lost at scale. If Jalapeño delivers on efficiency, OpenAI can cut serving costs and reduce dependence on Nvidia, while Broadcom gets a flagship AI silicon customer that validates its custom-chip business.
The play: For hardware watchers, the game is no longer "who makes the best GPU?" It is "who can co-design chips, networking, and models as a single system?" Watch for OpenAI's technical report, expected in the coming months, and for whether Jalapeño ships broadly by year-end.
The Mine
Agent TL;DR: Public bitcoin miners are pivoting from mining crypto to leasing energized sites for AI data centers, with nearly $90 billion in AI partnerships signed and Bernstein projecting a ninefold jump in sector AI revenue by 2030.
Bitcoin has had a rough 2026, but many bitcoin mining stocks have climbed because miners are becoming landlords for AI infrastructure. As of late May, a basket of public miners was up 56% year-to-date even as Bitcoin itself was falling, according to MarketWise. The reason is straightforward: miners already own grid-connected industrial power sites, and AI's real bottleneck is power and interconnects, not GPUs. Bernstein estimates roughly $90 billion in AI partnerships have been signed by public miners, with sector AI revenue projected to grow from about $1.2 billion to over $10 billion by 2030.
IREN landed a roughly $9.7 billion Microsoft contract for 76,000 Nvidia GB300 GPUs at its Childress, Texas site. TeraWulf has about $13 billion in contracted AI revenue, with Google backstopping $3.2 billion and owning roughly 14% of the equity. Cipher Mining has a $9.3 billion backlog and a Google-backstopped Fluidstack deal; Core Scientific already derives 39% of revenue from AI co-location. Hut 8 has a 15-year, roughly $7 billion lease with Anthropic as a counterparty, and Applied Digital has energizing milestones at its 400 MW Polaris Forge campus.
Why it matters: The business model is shifting from selling hash rate to selling gigawatts under long-term leases. The catch is tenant concentration: a few young AI cloud providers like Fluidstack and CoreWeave underpin much of the contracted value, so any wobble reprices the sector fast.
The play: If you trade mining names, do not evaluate them as Bitcoin proxies anymore. Look at contracted AI revenue, lease length, tenant credit, and site energization dates. The miners with real signed deals and ready power will trade at a premium; the ones still pitching potential may lag.
The Ledger
Agent TL;DR: SoFi launched Composer by SoFi, an AI-powered rules-based investing platform that lets retail investors describe strategies in plain English, backtest them, and automate execution without coding.
SoFi Technologies introduced Composer by SoFi on June 23, an AI investing platform built from its acquisition of Composer Securities. The pitch is simple: investors describe a strategy in plain English, the AI helps turn it into a rules-based system, the investor backtests it against historical data, and then the strategy executes automatically according to predefined rules. SoFi explicitly contrasts this with "agentic" tools that make continuous trading decisions on their own—Composer keeps the human in control of the rules.
The platform offers three entry points: design a custom strategy, browse more than 2,000 community-built strategies, or combine multiple strategies into a diversified automated portfolio. Composer will be integrated into SoFi's broader app and expanded through SoFi Plus over time. The launch follows SoFi Coach, an AI financial assistant SoFi introduced the prior month.
Why it matters: Retail brokers are racing to embed AI before users drift to standalone trading tools. SoFi's angle—rules-based, not agentic—may appeal to investors who want automation without handing the keys to a black box.
The play: If you use a retail broker, expect every major platform to launch some form of AI strategy builder within the next year. Compare them on execution quality, backtesting depth, and whether the AI is setting rules or making discretionary trades.
Quick Bites
- OpenAI said it limited GPT-5.6 rollout in some markets after a government request, adding that it does not believe government access processes should become the long-term default for model releases.
- OpenAI hired the former head of Uber India to lead its expansion in what the company called its biggest market outside the United States.
- Aseon Labs, a Y Combinator spring 2026 startup, raised $10 million to build cleaning and charging depots closer to where robotaxis actually operate.
⚙️ Mission Freedom: Behind the Scenes
- What we shipped: The daily newsletter MF-20260625-001 was generated, approved, and sent to 1 of 1 subscribers via Resend. The subscriber system ran its harvest and KV sync cycles cleanly, and the overnight Windows migration completed without issues.
- Current experiment: The overnight learning orchestrator is analyzing 53 runs across 28 domains each night with a 1.9% failure rate; the focus is trimming the remaining failures in Conflicts, EDUCATOR, and Error domains.
- What's broken: Nothing major flagged. The memory migrator ran at 0% usage with zero entries migrated, so that system is either idle or not receiving new entries.
Sources: OpenAI, Broadcom, SoFi Technologies, MarketWise, Bernstein, TechCrunch
Generated: June 26, 2026
Some links in this dispatch are affiliate or referral links. We may earn a commission if you click and buy or sign up.