Thursday, July 9, 2026

The July 1 and July 2 editions covered Meta's compute-sale exploration; today the story is scale-up, with Meta reportedly planning to roughly double compute capacity in 2027, buy-side capex estimates of $200–250 billion, compute deals with CoreWeave, Google, and Oracle, and SemiAnalysis flagging the ramp as aggressive alongside the $14.3 billion Scale AI talent acquisition. That demand is pulling adjacent assets into play — Lancium, the power developer behind OpenAI and Oracle's Stargate campus in Texas, is in minority stake-sale talks with Nvidia among interested parties, while Marvell completed its Celestial AI acquisition and Credo bought Hyperlume and DustPhotonics — so the monitorables are whether official guidance confirms the buy-side capex range and how much of the doubling is sourced from third-party clouds versus owned build.

Across the past two editions we covered SK Hynix's $28 billion Nasdaq ADR plan and roughly $7 billion of anchor indications; today the order book is reported near $170 billion — about six times the deal — amid debate over whether the ADR structure manufactures a valuation premium over the home listing. With Micron's record results, a persistent HBM shortage, and a newly quantified $30 billion Apple-Broadcom agreement reinforcing the demand backdrop, the open questions are where the ADR ultimately prices relative to that debated premium and whether the follow-on pipeline — Rebellions targeting a 2027 Korea IPO and ScaleFlux planning a fall listing — draws comparable demand.

Yesterday's edition covered the first AI-driven SaaS write-downs; today the displacement reaches a named global enterprise, with Sanofi replacing ServiceNow with an in-house Claude Code-based agent, while Mayfield finds 42% of Fortune 50–Global 2000 tech leaders running agentic AI in production and the generative AI economy at $110 billion in annual sales. The counter-evidence is equally specific — Fullcast measures sales efficiency down 28% and nearly 90% of firms report no measurable productivity gains — and pricing friction sharpened as Palantir CEO Alex Karp called token-based AI pricing 'completely wrong' while Simon Taylor of the Fintech Brainfood newsletter argued heavy token spend looks wasteful until quality effects justify the scale; the variable to watch is whether measured productivity begins to track deployment rates.

The June 29 edition covered GPT-5.6 Sol's release to a pre-approved allowlist; today OpenAI shipped the full family — Sol, Terra, and Luna — plus the full-duplex voice model GPT-Live, SpaceXAI and Cursor jointly launched Grok 4.5 ahead of SpaceX's planned $60 billion Cursor acquisition, and Sol now ranks second to Anthropic's Claude Fable 5 at roughly one-third the cost. With AWS and Google reportedly copying Fable 5's architecture and Rakuten reporting a 79% time-to-market reduction using Claude Code, a close substitute at a third of the price puts the benchmark leader's premium under observable pressure; the monitorable is whether Fable 5's pricing or usage share shifts in response.

The July 2 and July 3 editions covered the overproduction short thesis and private-fund redemption stress; today's markers are pricing and terms — Bank of America lent OpenAI $520 million, a shift from the zero-debt disclosure covered June 22, BlackRock TCP markdowns flagged private-credit stress, Wonder is raising at a $9 billion valuation with an investor IPO ratchet and a planned $200 million personal contribution from Wonder founder Marc Lore, and Blue Origin sought $10 billion at a $130 billion valuation. Michael Burry, author of Cassandra Unchained, questions whether hyperscaler capex can recover its cost while bulls cite Jevons Paradox and falling inference costs expanding demand; the live gauge is GPU availability, since a move from sold-out rationing toward open availability would be the earliest observable change in that debate.

• OpenAI launched GPT-5.6 Sol, Terra, and Luna plus full-duplex voice model GPT-Live; Sol ranked second to Anthropic's Claude Fable 5 at one-third the cost.?
• SpaceXAI and Cursor jointly launched Grok 4.5; Cursor is building a general work agent ahead of SpaceX's planned $60 billion acquisition.?
• Anthropic's Claude Fable 5 architecture drew rapid copying by AWS and Google; Rakuten reported 79% time-to-market reduction using Claude Code.?
• Meta launched Muse Image and Muse Video generative tools; startup PrismML claims it ran a 27-billion-parameter Qwen model on an iPhone 17 Pro.?

• WSTS projects the 2026 semiconductor market at $1.51 trillion, up ~90%; DRAM contract prices surged 93–98% quarter-on-quarter in Q1 2026.?
• SK Hynix's $28 billion Nasdaq ADR offering drew roughly $170 billion in orders amid debate over a manufactured ADR valuation premium.?
• SambaNova raised $1 billion at an $11 billion valuation; Rebellions targets a 2027 Korea IPO; ScaleFlux plans a fall IPO.?
• South Korea pledged KRW 800 trillion for semiconductors; Micron posted record results amid a persistent HBM shortage; Apple signed a $30 billion Broadcom deal.?

• Beijing authorized Alibaba, ByteDance, and DeepSeek to buy Nvidia H200 chips in July 2026, capped near 200,000 units with security conditions.?
• Nvidia had paused H200 China production in January after blocked shipments, then resumed sales in March during a brief approval window.?
• Access is limited to the largest Chinese tech firms through government-supervised channels, muting commercial impact versus a full market opening.?
• Nvidia is reportedly pursuing partnerships with rival chip companies as a hedge against rising AI accelerator competition.?

• Meta plans to roughly double compute capacity in 2027; buy-side capex estimates range $200–250 billion, alongside compute deals with CoreWeave, Google, and Oracle.?
• Lancium, the power developer behind OpenAI and Oracle's Stargate campus in Texas, is in minority stake-sale talks, with Nvidia among interested parties.?
• Marvell completed its Celestial AI acquisition and Credo bought Hyperlume and DustPhotonics as MicroLED optical interconnect attracted listed-company clustering.?
• Rhodium Group found off-grid gas data centers fail a climate investor test; Renew Home, Sunrun, and Tesla pursue a 16 GW residential flexibility program.?

• Scott Alexander of Astral Codex Ten proposed 'Plan A': delay superintelligence until 2040 with mutual compute limits and physical data-center verification.?
• Mieke Eoyang, former US Deputy Assistant Secretary of Defense for Cyber Policy, warned frontier models on Chinese hardware raise serious cyber risks.?
• Anthropic rejected Chinese government claims of a Claude Code security backdoor; Redwood Research argued current models show misalignment signs.?
• SemiAnalysis flagged Meta's aggressive compute ramp and $14.3 billion Scale AI talent acquisition; Dina Powell McCormick joined Meta as President and Vice Chairman.?

• The generative AI economy reached $110 billion in annual sales; Mayfield found 42% of Fortune 50–Global 2000 tech leaders run agentic AI in production.?
• Sanofi is replacing ServiceNow with an in-house Claude Code-based agent; Prime Intellect raised $130 million for custom enterprise agents.?
• Anthropic's Claude caught two $18,000 errors behind a $240 variance; Fintech Brainfood's Simon Taylor argued 'tokenmaxxing' looks wasteful until quality effects justify scale.?
• Fullcast found sales efficiency down 28%; nearly 90% of firms report no measurable AI productivity gains; Microsoft CEO Satya Nadella emphasized core competencies.?

• The AI capex boom exceeds railway and dotcom cycles, but GPUs are sold out and rationed rather than freely available to retail buyers.?
• Michael Burry, author of Cassandra Unchained, questioned hyperscaler capex recovery; bulls cite Jevons Paradox and falling inference costs expanding demand.?
• Wonder is raising at a $9 billion valuation with an investor IPO ratchet; founder Marc Lore plans a $200 million personal contribution.?
• Bank of America lent OpenAI $520 million; Blue Origin sought $10 billion at a $130 billion valuation; BlackRock TCP markdowns flagged private-credit stress.?

• Firebird plans a 100MW Armenian data center with ~50,000 Nvidia Blackwell GPUs, with US export approval already granted.?
• Rev Lebaredian, Nvidia vice president of Omniverse and simulation technology, said Armenia could rank among the top five countries by GPU count.?
• US Secretary of State Marco Rubio and Armenian Foreign Minister Ararat Mirzoyan signed a Comprehensive Strategic Partnership Charter, critical-minerals MOU, and TRIPP Framework.?
• OpenAI partnered with Armenia on AI-native education; Microchip Technology received a US export license to expand FPGA development there.?

• The x402 protocol logged 16 million transactions and $851,000 volume in 30 days across 70,000 buyers and 36,000 sellers.?
• Apify and Coinbase expanded x402-accessible tools past 20,000; Cloudflare launched a Monetization Gateway charging via x402.?
• Over $7.2 billion migrated from LayerZero to Chainlink's CCIP as Mantle joined; Aave Labs introduced fixed-rate Stable Vaults.?
• Hyperliquid Policy Center and Phantom asked the CFTC to confirm onchain code is not a financial service.?

• Palantir CEO Alex Karp called token-based AI pricing 'completely wrong' while Palantir launched a secure US-agency product using Nvidia Nemotron models.?
• Japan's Prime Minister Sanae Takaichi won a February 2026 supermajority as immigration debate intensified; zero immigration would accelerate labor-force contraction.?
• Tesla eliminated safety drivers from Miami robotaxi operations, moving toward fully driverless commercial deployment.?
• Millennium Management's index-rebalancing desks made $3.7 billion in June; the Hyperliquid-linked PURR fund's NAV premium averaged 1.04 since December.?