RipperMercs · github-actions · Jun 23, 2026 · Jun 23, 2026
diff --git a/public/llms.txt b/public/llms.txt
@@ -592,7 +592,8 @@ Substrate changelog (human): https://tensorfeed.ai/substrate (model lifecycle, M
 - [Has frontier training compute slowed](https://tensorfeed.ai/verdicts/compute-growth-slowdown): TF Verdict: not at the ceiling, frontier training compute is still climbing roughly 4 to 5x per year, but the curve is bending below it as total per-flagship training compute flattens and labs reroute spend into reinforcement learning. TF Verdict, May 29, 2026.
 - [Should you trust AI-found CVEs](https://tensorfeed.ai/verdicts/trust-ai-found-cves): TF Verdict: no by default; trust the AI pipeline that ships a working reproduction and a human gate, and treat any unreviewed bulk AI finding as an unconfirmed lead, not a CVE, until someone reproduces it. TF Verdict, May 29, 2026.
 - [Is the frontier premium worth it over open models](https://tensorfeed.ai/verdicts/frontier-premium-worth-it): TF Verdict: for most agent tasks, no; route default traffic to open weights at the inference floor and reserve the frontier premium for long-horizon agentic coding and high-stakes reasoning where a roughly 8-point benchmark gap compounds across a trajectory. TF Verdict, May 29, 2026.
-- [Originals](https://tensorfeed.ai/originals): Original editorial articles by TensorFeed (140 articles)
+- [Originals](https://tensorfeed.ai/originals): Original editorial articles by TensorFeed (141 articles)
+- [Reflection Pre-Bought $6.3 Billion of Colossus Compute Without a Shipped Model. The Open-Source Frontier Just Got a Procurement Story.](https://tensorfeed.ai/originals/reflection-ai-6b-colossus-open-frontier-compute): On June 22, 2026, Reflection AI signed with SpaceX for $150 million a month of Nvidia GB300 capacity at Colossus 2, starting July 1 and running through 2029, totaling roughly $6.3 billion. Reflection is a $25 billion open-source frontier lab with no publicly shipped model, founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, with Department of Energy Genesis Mission and Pentagon AI work already on the customer list. Read against SpaceX's prior Colossus commitments (Anthropic ~$45B, Google ~$30B, plus the Cursor acquisition), it is the third frontier-tier lease in seven months and the first one for a lab that has not yet released weights. Inside the per-GPU math, why Colossus is doing the Equinix move at the AI layer (neutral, multi-tenant gigawatts the hyperscalers cannot unbundle from a managed-service tax), the Pentagon-clearance angle that separates Reflection from DeepSeek and Z.ai, and three signposts in the next ninety days that decide whether $6.3B is a floor or a ceiling. The 90-day notice clause matters more than the headline number. Marcus Chen, June 23, 2026.
 - [China Drafted a $295 Billion State AI Grid. The Compute Race Now Runs on Two Different Rails.](https://tensorfeed.ai/originals/china-295b-state-ai-grid-sovereign-rail): Bloomberg surfaced China's National Development and Reform Commission blueprint for a 2 trillion yuan ($295B) five-year national AI compute network, financed by sovereign debt and ultra-long special government bonds, operated by China Mobile and China Telecom, and supplied 80 percent by domestic chipmakers led by Huawei. The grid is targeted to connect by 2028, and the procurement mandate excludes Nvidia and AMD by design. Read against Anthropic's $200B private commitment to Google TPU and the hyperscaler equity loop financing US frontier compute, the structural picture is two parallel rails financing the same scarcity with very different failure modes: the American rail is private, equity-backed, and demand-pull; the Chinese rail is sovereign, fiscal, and supply-push, with the operator layer consolidated inside the state telco duopoly. Inside the financing math, the Huawei HBM ceiling that decides whether 2028 is real or a slide, why state-directed buildout can internalize externalities the hyperscaler loop cannot, what multi-rail routing means for builders shipping into both markets, and three signposts in the next ninety days that convert the $295B planning number into a budget or back into a draft. Marcus Chen, June 22, 2026.
 - [Google Paid $2.7 Billion to Bring Shazeer Back. He Walked to OpenAI 22 Months Later. The Acqui-Hire Cliff Just Got a Price.](https://tensorfeed.ai/originals/shazeer-google-openai-acqui-hire-cliff): On June 18, 2026, Noam Shazeer, Google's VP of Engineering and co-lead of Gemini, told staff he was leaving for OpenAI. Twenty-two months earlier, in August 2024, Google paid roughly $2.7 billion in a CharacterAI licensing deal that was structurally an acqui-hire designed to keep him in the building. The retention clock just hit zero on the most expensive single engineer Google has ever bought back, and the destination is the rival walking into the IPO window with the most aggressive talent budget in the industry. The 2024 deal had the same shape as Microsoft-Inflection, Amazon-Adept, and Meta-Scale: a non-exclusive license dressed over a retention contract, engineered to slip past antitrust. The Shazeer departure is the first time the named principal has walked, and it sets a public price on the cliff that every other lab can now read. Inside the deal math, why 22 months is the cliff and not the contract, what it does to a Gemini 3.5 Pro launch that is already slipping, and what it costs OpenAI to make a hire this public 30 days after the $150M Partner Network move and 90 days after the $122B raise. Marcus Chen, June 21, 2026.
 - [OpenAI Put $150 Million Behind 300,000 Consultants. The Partner Network Is a Channel Moat Against Anthropic.](https://tensorfeed.ai/originals/openai-partner-network-150m-channel-moat): On June 14, 2026 OpenAI announced the OpenAI Partner Network, a $150 million channel program structured around Select, Advanced, and Elite tiers, with a target of 300,000 certified consultants by year end and launch partners including Accenture, BCG, McKinsey, Bain, PwC, Eliza, and Artium. Specializations cover Codex, cybersecurity, API, and agent transformation, and a Forward Deployed Experts pilot embeds partner practitioners alongside OpenAI engineers on Elite engagements. It is the second OpenAI implementation move in five weeks, after the $4 billion Deployment Company in May, and it lands 30 days after the Ramp AI Index put Anthropic ahead of OpenAI on enterprise spend at 41 percent of paying US businesses. When the model commoditizes, the value migrates to whoever owns the implementation layer, and OpenAI just bought a 300,000-strong consulting army whose comp plans are now structurally tilted toward recommending GPT-class models first. The channel is the moat; the Big Four pen is the new sales motion. The question for Anthropic is whether the Seoul-style sovereignty bundle and Claude Code's developer surface beat a Big-Four-led procurement check. Marcus Chen, June 20, 2026.