From e463bd405ba717c75760e0d877b2982e7d5f13f1 Mon Sep 17 00:00:00 2001 From: tensorfeed-factcheck-bot Date: Mon, 27 Jul 2026 16:17:17 +0000 Subject: [PATCH] factcheck: corrections to originals [skip-factcheck] --- public/llms.txt | 2 +- src/app/originals/amd-anthropic-2gw-mi450-fifth-vendor/page.tsx | 2 +- .../naver-nvidia-brookfield-10b-sovereign-ai-triangle/page.tsx | 2 +- src/lib/originals-directory.ts | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/public/llms.txt b/public/llms.txt index 04eccac1..8c45b80e 100644 --- a/public/llms.txt +++ b/public/llms.txt @@ -625,7 +625,7 @@ Substrate changelog (human): https://tensorfeed.ai/substrate (model lifecycle, M - [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 (184 articles) -- [NAVER, NVIDIA, and Brookfield Put $10 Billion Behind Korea's 200 Megawatt Sovereign AI Factory. The Financing Template Is the Story.](https://tensorfeed.ai/originals/naver-nvidia-brookfield-10b-sovereign-ai-triangle): On Friday, July 25, 2026, NAVER, NVIDIA, and Brookfield announced a $10 billion expansion of NAVER's GAK Sejong DSX AI factory from 55 megawatts to 200 megawatts by 2028, with NVIDIA writing $1B of equity into NAVER, Brookfield committing up to $9B on a non-binding term sheet as exclusive capital partner, and NAVER carrying the remainder and running the site on Blackwell today and Vera Rubin from 2027, with a stated long-term ambition of 1 gigawatt on the same Sejong campus. The megawatt count is not the story. The financing structure is. The vendor-equity loop (Google recycling $40B into Anthropic TPU offtake, AMD writing $5B to close 2GW of MI450, NVIDIA sending $40B toward OpenAI to anchor Vera Rubin) only closes when the customer is a US frontier lab with hyperscaler credit behind it, and NAVER is not that customer. The gap gets filled by Brookfield's $100 billion AI infrastructure fund, itself anchored by NVIDIA and KIA: NVIDIA writes a token equity check for alignment, Brookfield writes the majority of the bill against the physical asset, the customer signs a compute offtake that services the debt, the whole thing gets built without a hyperscaler landlord. Inside the numbers table, the two sovereign AI templates now in the field (Beijing burns political will and takes the yield hit on domestic silicon, Seoul keeps the frontier stack and dilutes to allied counterparties), what a repeatable NVIDIA plus project-finance placement mechanism does to the sovereign compute growth vector nobody had priced (Korea, Japan, Taiwan, Singapore, the Gulf, and slices of the EU), the concentration risk when the same NVIDIA-anchored Brookfield fund shows up on every allied deal, why the 2028 delivery window puts NAVER's take-or-pay math on the same 24-month clock as every other frontier commitment written this year, and why the 1 GW ceiling is the number the deal actually depends on. Three signposts: whether the Brookfield term sheet hardens into binding debt before year end, whether Japan or Taiwan announces a comparable NVIDIA plus project-finance deal before Q1 2027, and whether the AI FINRA gate ends up applying to sovereign AI factories that host US-designed model weights. Marcus Chen, July 27, 2026. +- [NAVER, NVIDIA, and Brookfield Put $10 Billion Behind Korea's 200 Megawatt Sovereign AI Factory. The Financing Template Is the Story.](https://tensorfeed.ai/originals/naver-nvidia-brookfield-10b-sovereign-ai-triangle): On Friday, July 25, 2026, NAVER, NVIDIA, and Brookfield announced a $10 billion expansion of NAVER's GAK Sejong DSX AI factory from 55 megawatts to 200 megawatts by 2028, with NVIDIA writing $1B of equity into NAVER, Brookfield committing up to $9B on a non-binding term sheet as exclusive capital partner, and NAVER carrying the remainder and running the site on Blackwell today and Vera Rubin from 2027, with a stated long-term ambition of 1 gigawatt on the same Sejong campus. The megawatt count is not the story. The financing structure is. The vendor-equity loop (Google recycling $40B into Anthropic TPU offtake, AMD writing $5B to close 2GW of MI450, NVIDIA sending $30B toward OpenAI to anchor Vera Rubin) only closes when the customer is a US frontier lab with hyperscaler credit behind it, and NAVER is not that customer. The gap gets filled by Brookfield's $100 billion AI infrastructure fund, itself anchored by NVIDIA and KIA: NVIDIA writes a token equity check for alignment, Brookfield writes the majority of the bill against the physical asset, the customer signs a compute offtake that services the debt, the whole thing gets built without a hyperscaler landlord. Inside the numbers table, the two sovereign AI templates now in the field (Beijing burns political will and takes the yield hit on domestic silicon, Seoul keeps the frontier stack and dilutes to allied counterparties), what a repeatable NVIDIA plus project-finance placement mechanism does to the sovereign compute growth vector nobody had priced (Korea, Japan, Taiwan, Singapore, the Gulf, and slices of the EU), the concentration risk when the same NVIDIA-anchored Brookfield fund shows up on every allied deal, why the 2028 delivery window puts NAVER's take-or-pay math on the same 24-month clock as every other frontier commitment written this year, and why the 1 GW ceiling is the number the deal actually depends on. Three signposts: whether the Brookfield term sheet hardens into binding debt before year end, whether Japan or Taiwan announces a comparable NVIDIA plus project-finance deal before Q1 2027, and whether the AI FINRA gate ends up applying to sovereign AI factories that host US-designed model weights. Marcus Chen, July 27, 2026. - [MCP Goes Stateless on Monday. The Session Handshake Is Gone, and So Is the Reason Servers Were Hard to Run.](https://tensorfeed.ai/originals/mcp-stateless-monday-session-handshake-gone): The Model Context Protocol ships its biggest revision since launch on Tuesday, July 28, 2026, with the 2026-07-28 release candidate signed off by lead maintainers David Soria Parra and Den Delimarsky. The change list rearranges the entire runtime: the initialize / initialized handshake is removed, the Mcp-Session-Id header is removed, and any request can hit any server instance because client information now travels via a _meta field on every call. Extensions become first-class with reverse-DNS identifiers, capability negotiation, and independent version cadence, and the first two official extensions ship alongside the core. MCP Apps lets a server ship interactive HTML that the host renders in a sandboxed iframe with templates prefetched and security-reviewed, so a tool call can become a small application surface instead of a JSON round trip. Tasks graduates out of the experimental core into SEP-2663, blocking tasks/result is replaced by polling via tasks/get, tasks/update, and tasks/cancel, session-bound tasks/list is removed outright, and servers return task handles from tools/call that the model reasons about. Six authorization proposals harden the spec against real OAuth 2.0 and OpenID Connect deployments (RFC 9207 iss validation, application_type on registration, refresh token guidance), with Enterprise-Managed Authorization stable on July 6, 2026 and Okta named as the reference identity provider alongside Anthropic's Claude, Claude Code, Cowork, plus VS Code as reference clients. Roots, Sampling, and Logging get deprecated in the core, and a formal Active / Deprecated / Removed policy with a 12-month minimum window replaces silent breaking changes between spec revisions. Inside the changes table, why the stateless bet is the whole story (a request behind a normal load balancer, autoscaling on horizontal replicas, serverless without a shared session store), what breaks (initialize-first servers, session-bound task queues, tasks/list-driven UIs), and the migration pattern of explicit handles instead of transport-hidden state. The edge angle: the stateless move is the change that finally lets MCP land clean on Cloudflare Workers and AWS Lambda cold starts, which is exactly the mental model x402 already assumed when the monetization gateways shipped two weeks ago. Three signposts: whether Anthropic ships MCP Apps in Claude Desktop and Claude Code by Q4, whether OpenAI adopts the auth hardening for the ChatGPT MCP surface, and whether the first paid x402-metered MCP Apps land on any hosted registry before Q3 close. Adrian Vale, July 26, 2026. - [Google Cloud Booked $514 Billion in Backlog. Q2 Was the First Quarter TPUs Shipped Into Customer Data Centers.](https://tensorfeed.ai/originals/google-cloud-514b-backlog-tpu-in-customer-dc): Alphabet reported Q2 2026 after the close on Tuesday, July 22. Google Cloud grew 82 percent YoY to $24.8B with a 35.6 percent operating margin ($8.81B, up from $2.83B), the cloud backlog jumped $54B in a single quarter to $514B (up 385 percent YoY, roughly $257B set to convert inside 24 months), Q2 CapEx nearly doubled to $44.9B, full-year 2026 CapEx guidance was raised to $195B to $205B, and free cash flow printed roughly negative $5.9B. Most wires ran the after-hours stock drop as the story. The story CFO Anat Ashkenazi actually gave on the call is one sentence: Google recognized revenues from TPU system sales delivered to customer data centers for the first time in Q2. That is a chip line, not a cloud line, and it puts Alphabet inside Nvidia's business model on top of the cloud business already running against AWS and Azure. Inside the full numbers table, the merchant-silicon versus bundled-sale question the disclosure did not answer, why the backlog print is the first hard data point for the bull side of the CapEx debate, and what a working TPU-in-customer-rack product does to Nvidia and Broadcom. Three signposts: whether Q3 breaks out TPU-system revenue as a named line, whether backlog crosses $600B, and what Nvidia says on the late-August call about custom-silicon customer mix. Marcus Chen, July 25, 2026. - [Anthropic Shipped Opus 5 at the Old Opus Price. It Beats Fable 5 on Most Rows for Half the Money.](https://tensorfeed.ai/originals/claude-opus-5-same-price-half-of-fable): Claude Opus 5 landed July 24, 2026 at $5 per million input tokens and $25 per million output, exactly what Opus 4.8 costs and exactly half of Fable 5, so every benchmark gain in the release is free in dollar terms. Against Opus 4.8 the gains are lopsided: ARC-AGI-3 goes 1.5 to 30.2 percent (a capability the previous model did not have, on a benchmark built to resist memorization), computer use on OSWorld 2.0 climbs 55.7 to 70.6, agentic terminal coding on Frontier-Bench doubles 21.1 to 43.3, and GDPval-AA v2 knowledge work goes 1593 to 1861. Against Fable 5 the picture is awkward for Anthropic: across the eleven rows where both report a directly comparable number, Opus 5 takes seven and Fable 5 four, with two of Fable's four inside two tenths of a point, yet Anthropic still calls Fable 5 the top capability tier at double the price. Two chart rows are excluded as non-comparable because the Fable column reports a Mythos 5 ceiling from the safeguards-lifted Project Glasswing variant. The upgrade is free but the port is not: thinking is on by default so a request that omitted the parameter now spends more and can truncate against a tight max_tokens, disabling thinking above high effort returns a 400 where 4.8 accepted it, the prompt cache minimum halves to 512 tokens, Opus 5 uses a separate rate-limit pool, and Priority Tier does not cover it. Three signposts: whether ARC-AGI-3 survives independent replication, whether Fable 5 gets repriced, and why no SWE-bench Verified number appears on a coding-focused launch table. Adrian Vale, July 24, 2026. diff --git a/src/app/originals/amd-anthropic-2gw-mi450-fifth-vendor/page.tsx b/src/app/originals/amd-anthropic-2gw-mi450-fifth-vendor/page.tsx index f1cd8b22..481c3d5e 100644 --- a/src/app/originals/amd-anthropic-2gw-mi450-fifth-vendor/page.tsx +++ b/src/app/originals/amd-anthropic-2gw-mi450-fifth-vendor/page.tsx @@ -165,7 +165,7 @@ export default function AmdAnthropic2gwMi450FifthVendorPage() {