The Bleeding Edge

// Episode W27 · 2026-06-25 to 2026-07-02

This was the week the AI fight moved from software to silicon

This was the week the AI fight moved from software to silicon. For three years the frontier was measured in benchmarks; this week it was measured in memory chips. The Financial Times reported Apple is lobbying the Trump administration for permission to buy memory chips from black…

The Bleeding Edge — Episode Briefing W27

Date range: 2026-06-25 to 2026-07-02 (Europe/Madrid)

Headline of the Week

This was the week the AI fight moved from software to silicon. For three years the frontier was measured in benchmarks; this week it was measured in memory chips. The Financial Times reported Apple is lobbying the Trump administration for permission to buy memory chips from blacklisted Chinese suppliers, because it can't get enough anywhere else — and the labs kept buying their way down the stack: Micron signed an HBM-supply-plus-investment deal with Anthropic, OpenAI detailed its first custom chip, and Dell put NVIDIA's GB10 on a desktop. It all played out against a hardening trade backdrop — on 22 June China barred 56 US firms (mostly defense contractors and rare-earth miners) in retaliation for a Pentagon military blacklist, a reminder that chip and materials supply is now geopolitics. And the model cadence didn't so much as blink — OpenAI previewed GPT-5.6 "Sol," DeepSeek shipped a decoding trick that runs V4 up to 85% faster, and Anthropic's Fable 5 / Mythos — the models Washington yanked offline worldwide last month — came back. The throughline: the constraint on AI is no longer ideas or even models. It's who can physically source the memory, the fab capacity, and the export license — and that is now a matter of statecraft.

Top 5

  1. The memory squeeze becomes AI's binding constraint. The FT reported Apple is lobbying the Trump administration to buy memory chips from blacklisted Chinese suppliers to cover a shortfall, while Micron signed an HBM-supply-plus-investment deal with Anthropic and OpenAI detailed its first custom chip — labs and device makers racing to lock up HBM and DRAM. It played out against a harder trade backdrop: on 22 June China barred 56 US firms — mostly defense contractors and rare-earth miners — retaliating for a Pentagon military blacklist (not AI export controls specifically). Why it matters: the constraint on AI has shifted from models to the physical substrate — memory, fab capacity, export licenses — and sourcing it is now foreign policy. Corroborated Sources: Financial Times, CNBC — China curbs 56 US firms.

  2. OpenAI launches Daybreak — GPT-5.5-Cyber and "Patch the Planet," AI that patches software vulnerabilities at scale. OpenAI unveiled Daybreak, a security initiative built on a specialized GPT-5.5-Cyber model, plus a "Patch the Planet" push aimed at automated discovery and remediation of vulnerabilities. Why it matters: this productizes — as defense — the exact "read a codebase and fix its flaws" capability that got Anthropic's Fable 5 pulled offline in W25, reframing frontier offensive-security as an enterprise patching service. Unverified — single-source (newsletter digest); no primary OpenAI post independently confirmed. Sources: The Creators AI.

  3. The model wave rolls on: GPT-5.6 "Sol," DeepSeek-V4 + DSpark, and Fable 5 / Mythos back from the dead. OpenAI is previewing GPT-5.6 Sol, a release aimed at pushing reasoning further; DeepSeek shipped DSpark, a speculative-decoding method it says runs DeepSeek-V4 57–85% faster than its prior MTP-1; and Anthropic's Fable 5 / Mythos — pulled worldwide by US Commerce in June — are back online ("Mythos ban lifted"). Why it matters: a full-stack model refresh from three labs in one week says the release cadence is accelerating, not cooling, even under an export-control regime. Corroborated (aggregate + DeepSeek primary); GPT-5.6 Sol specifics are single-source. Sources: AI Search — "GPT-5.6, IBM chip, Ornith 1.0, Seed 2.1, Mythos ban lifted", MarkTechPost — DeepSeek DSpark.

  4. John Jumper — Nobel laureate and AlphaFold creator — leaves Google DeepMind for Anthropic. Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, is reportedly departing DeepMind to join Anthropic. Why it matters: it's the highest-profile science-AI defection to date and a signal that the talent war is now about who owns the scientific-discovery frontier, not just chat and code — and DeepMind just lost its most decorated researcher to a direct rival. Unverified — single-source (newsletter digest); no primary confirmation from either lab in-corpus. Sources: The Creators AI.

  5. The labs buy their way into silicon: Micron–Anthropic HBM deal, OpenAI's first chip, Dell's GB10 desktop. Micron signed an AI-infrastructure agreement with Anthropic pairing HBM memory supply with a Series H investment; OpenAI detailed its first custom AI chip; and Dell shipped a workstation built on NVIDIA's GB10. Why it matters: with memory now a geopolitical chokepoint (see #1), every major lab is vertically integrating into hardware — and Micron becoming both Anthropic's supplier and investor echoes the same supplier-as-stakeholder entanglement that defined the Fable 5 saga. Corroborated (Micron–Anthropic) / Unverified (OpenAI chip specifics). Sources: The Creators AI, The Neuron.

Categorised News

Market Cap / Valuation

  • SoftBank's $500 billion question. Masayoshi Son reportedly told shareholders that SoftBank's massive AI-infrastructure bet (the Stargate-scale commitments) rests on assumptions he can't fully prove out — a rare public hedge from AI's most aggressive financier. Read it as the first crack of doubt from inside the money behind the buildout. Unverified (single-source analysis piece); framing is Inference. Source: Product Market Fit.
  • Bending Spoons IPOs. The Italian app conglomerate behind AOL, Vimeo, and Eventbrite went public; the NYT notes its playbook of acquiring mature software brands and then cutting staff hard. A tell for how AI-era efficiency logic is being applied to legacy consumer software. Corroborated Source: New York Times.

Frontier & Big Tech

  • Fable 5 / Mythos ban lifted. Anthropic's flagship models — forced offline globally by US Commerce three days after their June launch — are back in service. The fastest reversal-to-restoration of a frontier model to date, and a marker that the export-control regime is still being negotiated in real time. Corroborated (two aggregators + W25 continuity). Sources: AI Search, The Neuron.
  • Getty Images + OpenAI strike a multi-year licensing deal. Licensed Getty photos are now available inside ChatGPT — a clean, paid alternative to scraped training data and a template for image-rights licensing. Unverified (single-source). Source: The Creators AI.
  • The rest of the model roundup: IBM's new AI chip, ByteDance's Seed 2.1, Ornith 1.0, and an "Aleph brain scan." A dense week of releases and interpretability work across the field, per AI Search's digest — individually thin, collectively another data point that the release pace is relentless. Unverified (single aggregator). Source: AI Search.

Apps / Dev Tools / Platforms

  • Meta open-sources Astryx. A React design system with 90+ components, an MCP server, and a CLI built for AI coding agents — Meta shipping the scaffolding so agents (not just humans) can assemble front-ends. A strong signal that "design systems for agents" is becoming its own category. Corroborated Source: MarkTechPost.
  • EverOS — open-source agent memory runtime. Markdown-first, hybrid BM25 + vector retrieval, with self-evolving "skills." Directly relevant to anyone (us included) building durable agent memory without a database. [Corroborated as announced] Source: MarkTechPost.
  • Claude Tag. Anthropic shipped a configurable team agent for Slack — setup, permissions, and controls for a shared workspace assistant. Unverified (single-source). Source: The Creators AI.

Infrastructure & Ecosystem

  • DeepSeek DSpark. A speculative-decoding technique DeepSeek says runs V4 57–85% faster than its previous MTP-1 approach — pure throughput, no new parameters. Treat the exact percentages as vendor-reported. [Corroborated as announced] (vendor benchmark). Source: MarkTechPost.
  • NVIDIA BioNeMo Agent Toolkit. Wraps biomolecular models as callable skills for drug-discovery agents — the "agents calling scientific tools" pattern arriving in life sciences. [Corroborated as announced] Source: MarkTechPost.
  • Hugging Face + Cerebras deepened an inference partnership, pushing high-token-per-second serving to the HF ecosystem. Unverified (single-source). Source: The Neuron.

AI in Consumer Hardware

  • Dell Pro Max with GB10. Dell productized NVIDIA's GB10 into a desktop AI workstation — local frontier-class inference on your desk, and another beneficiary of the shift toward on-device compute as cloud costs and chip politics bite. Unverified (single-source). Source: The Neuron.

Prompting Skill of the Week

Name: Rubric-First Prompting (a.k.a. "grade the test before you take it") Best for: any task where "good" is subjective or high-stakes — drafting a board memo, extracting fields from a contract, writing a customer email — where a one-shot answer looks confident but drifts from what you actually needed.

The technique: before the model does the work, make it write (or accept) the rubric it will be judged on, then produce the answer, then score its own answer against that rubric and revise.

Steps:

  1. State the job and the audience — "Draft a 200-word update to a non-technical board on our AI pilot."
  2. Ask for the rubric first — "Before writing anything, list the 5 criteria a great version must meet (e.g., no jargon, one clear ask, quantified result)."
  3. Approve or edit the rubric — add the criterion the model missed; delete the fluff. This is where your judgment goes in.
  4. Now produce the answer — "Write the update to satisfy every criterion above."
  5. Self-score — "Score the draft 1–5 on each criterion and quote the exact line that earns or loses each point."
  6. Revise on the weakest score only — cheap, targeted, and it stops the model from rewriting the good parts.

Example prompt (one message):

"We'll do this in three passes. Pass 1: list the 5 criteria a great answer must meet — don't write the answer yet. Pass 2: write the answer to hit all 5. Pass 3: score yourself 1–5 per criterion, quote the line for each score, then rewrite only the lowest-scoring part. The task: [task]."

Common failure + fix: the model writes a flattering rubric that its answer conveniently passes. Fix: supply one or two non-negotiable criteria yourself ("must fit in 200 words," "must name a specific dollar figure or say 'not yet measured'") so it can't grade on a curve.

New AI Tools

  1. xAI Voice Agent Builder — a no-code builder for Grok Voice agents aimed at support, sales, and scheduling calls. Point-and-configure a voice agent that answers phones, qualifies leads, or books appointments without writing code. Who for: operations and CX leaders who want a phone-answering agent live this week, not a six-month integration project. Sources: The Neuron.
  2. Mistral OCR 4 — Mistral turned document extraction into a full enterprise AI play: OCR that reads complex, multi-format documents and hands back structured data, positioned as an on-ramp to Mistral's broader stack. Who for: any team drowning in invoices, contracts, or forms that still get keyed in by hand. Sources: The Creators AI.
  3. Liquid AI LFM2.5-230M — a 230M-parameter model pretrained on 19T tokens that Liquid says beats models 4× its size on data extraction and "runs anywhere," including on edge hardware. Who for: builders who need private, cheap, offline extraction on a laptop or device rather than a cloud API call. Sources: MarkTechPost.

AI Personality of the Week

John Jumper. Jumper is the DeepMind scientist who led AlphaFold, the system that cracked protein-structure prediction and, in doing so, reorganized computational biology — work that earned him a share of the 2024 Nobel Prize in Chemistry. This week he is reported to be leaving Google DeepMind for Anthropic, a move that would make him the most decorated researcher yet to switch labs and would plant a Nobel-caliber science pedigree squarely inside Anthropic. Where a scientist of Jumper's stature chooses to work on "AI for discovery" is a leading indicator of where the next AlphaFold-scale breakthrough gets built — and, if confirmed, DeepMind just handed that indicator to a competitor. Unverified — single-source; awaiting primary confirmation. Source: The Creators AI.

Catch-All

Michael Burry is shorting the AI buildout. Burry — the Big Short investor — disclosed short positions against Caterpillar, a bet that reads as skepticism toward the data-center construction boom, while short-sellers separately piled into SpaceX ahead of and after its blockbuster IPO (a trade Reuters notes has already cost them as the stock climbed). It's the clearest sign that the smart-money contrarians think at least part of the AI-infrastructure story is priced for perfection — worth watching as a sentiment counterweight to SoftBank's half-trillion-dollar conviction. Corroborated Sources: Reuters, Capital Brief.

Show Notes (bullets only)

  • 🇨🇳 China blacklists 56 US companies (24 June) in retaliation for AI export controls — the trade war goes kinetic.
  • 🍎 Apple lobbies the Trump admin to buy memory chips from blacklisted Chinese suppliers (per the FT) — the shortage is real.
  • 🛡️ OpenAI launches Daybreak — GPT-5.5-Cyber + "Patch the Planet," AI-powered vulnerability patching at scale. Unverified
  • 🧠 Model wave: OpenAI previews GPT-5.6 "Sol," DeepSeek's DSpark runs V4 up to 85% faster, and Fable 5 / Mythos are back online after last month's global ban.
  • 🧬 John Jumper (AlphaFold, 2024 Nobel) reportedly leaves DeepMind for Anthropic. Unverified
  • 💾 Micron–Anthropic sign an HBM-supply + Series-H-investment deal; OpenAI details its first chip; Dell ships a GB10 desktop.
  • 📸 Getty + OpenAI license real photos into ChatGPT; Mistral OCR 4 turns document extraction into an enterprise platform.
  • 🧩 Meta open-sources Astryx — a React design system with an MCP server and CLI built for AI coding agents.
  • 🗣️ xAI Voice Agent Builder ships no-code Grok Voice agents for support, sales, and scheduling.
  • 💸 SoftBank's $500B question: Son hedges on the AI-infra bet; Michael Burry shorts Caterpillar and shorts pile into SpaceX.
  • 🐣 Liquid AI LFM2.5-230M — a 230M model that beats models 4× its size on extraction and runs on the edge.

Weekly Patterns (Inference)

  1. The binding constraint moved from models to memory. Inference Every marquee story — the China blacklist, Apple's lobbying, the Micron–Anthropic deal, OpenAI's chip — is about physical supply, not capability. The bottleneck is now HBM, fab capacity, and export licenses.
  2. Supplier and stakeholder keep collapsing into one entity. Inference Micron becomes Anthropic's memory supplier and investor, the same entanglement that let Amazon (Anthropic's backer) trigger the Fable 5 shutdown last month. The neat org chart of who-sells-to-whom is dissolving.
  3. Offensive-security capability is being relabeled as a product. Inference The "AI reads code and fixes flaws" ability that got Fable 5 banned in W25 is, weeks later, OpenAI's "Patch the Planet." The same capability is a liability or a feature depending on who ships it and how they frame it.
  4. Regulation didn't stop the cadence — it rerouted it. Inference Fable 5 came back, GPT-5.6 previewed, DeepSeek optimized. Export control changed who can access models, not how fast they ship.
  5. "Agents calling tools" is spreading into every vertical. Inference Astryx (design), BioNeMo (drug discovery), EverOS (memory), xAI Voice (telephony) all package existing capability as callable skills for agents — the interface layer, not the model, is where this week's building happened.
  6. The contrarian money is starting to hedge. Inference Son's public doubt and Burry's Caterpillar short, in the same week, are the first coordinated signals that sophisticated investors see the AI-infrastructure buildout as at least partly overbuilt.
  7. Small-and-local is quietly the counter-trend to the chip crunch. Inference Liquid AI's 230M model and Dell's GB10 desktop point the same direction: when cloud compute is expensive and chips are political, cheap on-device inference gets more attractive.