NVIDIA signs $500B+ AI-infrastructure consortium with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR
NVIDIA Investor Relations · LiveMint · Aug 10–11
On August 10 NVIDIA signed MoUs with six of the world's largest asset managers to establish independent compute financing platforms that will mobilize over $500 billion of third-party capital for AI-infrastructure buildout over time. The structure is six independent vehicles, not a single fund, each backed by NVIDIA's ecosystem of frontier labs, enterprises, and AI clouds. Larry Fink (BlackRock): "This partnership deepens our relationship with NVIDIA, including through the AI Infrastructure Partnership." Jon Gray (Blackstone): "We continue to be enormous investors globally across the Nvidia ecosystem." The $500B is a private-credit answer to the "AI capex bubble" framing — frontier labs can now finance GPU purchases through consortium vehicles, converting capex into opex-like cash flows and pre-funding the next 18–24 months of frontier-API price decline.
Source →ai-infrastructurecapital-marketsnvidia
Meta ships Muse Glimmer (30B Apache-2.0 agentic model) and opens Muse Spark 1.2 weights — Zuckerberg lobbies Washington
Meta Research · CNBC · Tech Startups · Aug 10
Meta released Muse Glimmer on August 10 — a 30B dense multimodal agentic model under Apache 2.0 that runs on a single consumer GPU (≈18–20 GB at 4-bit) with 3.1× faster generation on RTX 5090 via DFlash speculative decoding. Beats Gemma4-31B and Qwen3.6-27B in class. In the same announcement, Zuckerberg opened the weights for Muse Spark 1.2 — the more powerful frontier model previously released as closed. Zuckerberg framed it explicitly as US-versus-China competitiveness: "US policy should actively support open-weight AI development to keep pace with the powerful open models coming from Chinese labs." Context: Tencent Hunyuan 3 has held #1 on OpenRouter by token usage since April 28; Alibaba Qwen3.8-Max shipped August 3; DeepSeek-V4-Flash-0731 shipped July 31; ByteDance Seedance 2.5 shipped August 8.
Source →open-weightsmetalocal-agent
OpenAI launches GPT-5.6-Cyber via Daybreak Red — cyber-permissive tier with 95% advanced-cyber completion
VentureBeat · The Hacker News · OpenAI Daybreak · Aug 10–11
OpenAI released GPT-5.6-Cyber on August 10, a fine-tuned variant of GPT-5.6 Sol trained specifically for advanced cybersecurity tasks — zero-day discovery, exploit-chain development — with reduced refusals on dual-use requests. Access is gated through Daybreak Red, a new tier for authorized vulnerability research, exploit validation, and security testing, joining Daybreak Blue (defensive). The model achieved 95% completion on advanced cybersecurity tasks and independently identified CVE-2026-15903 in the V8 JavaScript engine (an integer-conversion safety-check skip enabling out-of-bounds array index read/write); OpenAI disclosed to Google, which fixed it. GPT-5.6-Cyber joins Anthropic's Project Glasswing as the second frontier-lab cyber-tier — "responsible-disclosure gating" is now the standard pattern.
Source →openaicyber-tierdistribution-gating
Thinking Machines Inkling lands as the leading US open-weights model — 975B MoE at 41B active, Inkling-Small at 12B
Simon Willison · Artificial Analysis · Jul 15 / Jul 30
Thinking Machines Lab (Mira Murati) released Inkling on July 15 — a 975B-total / 41B-active MoE transformer, Apache 2.0, trained on 45T tokens of text, images, audio, and video. Debuted at 41 on the Artificial Analysis Intelligence Index — "the new leading U.S. open-weights model." Inkling-Small followed on July 30: 276B MoE with 12B active parameters that beats its larger sibling on several reasoning and agentic benchmarks at less than one-third the size. Tinker fine-tuning, controllable thinking cost, and encoder-free audio/vision are the differentiators. For TTL builders evaluating on-device agent stacks, Inkling-Small is the first sub-15B-active model within one point of the flagship on Intelligence Index — benchmark against it before committing to closed-API pricing.
Source →open-weightsthinking-machinesmoe
Tencent Q2 2026 earnings today — Hunyuan 3, WorkBuddy traction, and "substantial" AI capex increase
IG.com · Bloomberg Asia Trade · Aug 12
Tencent reports Q2 today; consensus expects revenue +9.6% YoY to RMB 202.2B, non-IFRS net profit +7.6% to RMB 67.8B. Hunyuan 3 Preview has held #1 on OpenRouter by token usage since April 28, and WorkBuddy is "currently the most widely used productivity AI agent in China." Tencent Cloud international business grew more than 40% YoY. Management has guided to a "substantial increase" in AI capex for H2 2026 — the most concrete forward-look for Chinese frontier-model capex. For TTL, the Q2 number is less important than the capex guide and WorkBuddy traction — WorkBuddy is the first major Chinese productivity agent with named enterprise distribution, and its token-monetization curve is the bellwether for Chinese agent-economics viability.
Source →tencenthunyuan-3workbuddy
Anthropic–Riot $9.1B 20-year compute deal locks 191 MW at Rockdale TX — Bitcoin miners are credible AI compute brokers
Bloomberg · Reuters · Aug 11
Anthropic locked 96 MW by December 2027 with the remaining 95 MW scaling through 2045 at Riot Platforms' Rockdale, Texas facility. RIOT jumped 25% overnight — the largest publicly disclosed AI-Bitcoin-miner compute lease. Combined with Anthropic's $13B Series F at $183B post-money and $5B+ run-rate revenue, the structural compute demand is now bankable. Bitcoin miners — with stranded power, cooling, and gigawatts of pipeline — are the cheapest source of incremental AI compute and the second source that no one was thinking about twelve months ago. For TTL's burst-compute needs (model training, large backfills, capacity spikes), Riot/Cipher/IREN/Marathon-class "ex-bitcoin-miner" GPU clouds are a credible 20–40%-cheaper alternative to AWS/Azure/GCP at peak.
Source →computebitcoin-minerslong-term-lease
US July CPI prints at 8:30 ET — Core CPI forecast 2.5% YoY, the macro pivot of the week
BLS · Financial Juice · Yahoo Finance · Aug 12
Consensus forecasts Core CPI YoY at 2.5% (vs. prior 2.6%) and headline YoY 3.4% (vs. 3.5%). Core MoM consensus +0.2%. Brent at $87.72 and WTI at $82.13 are adding roughly 10 bps of upward pressure on goods. 30-year Treasury yields hover near 20-year highs. A Core print at or above 2.6% would reopen the September FOMC rate-hike debate (Sept 15–16), which markets have largely written off. A hot print could move 10-year yields 15–25 bps and compress AI-stock multiples — watch SaaS (CRM, NOW, INTU) for a quick derate. If in-line, the AI capex narrative resumes dominance; the NVIDIA consortium announcement is the largest financing signal this year and the macro pivot determines how the market reads it.
Source →macrocpifomc
OpenAI S-1 expected within weeks — September IPO filing target, $730B private valuation
Build Fast With AI · Tech Journal · Aug 11
OpenAI's public S-1 prospectus is expected to appear on the SEC's EDGAR system in mid-to-late August 2026, roughly 15 days before any investor roadshow. The filing will disclose audited financials, the Microsoft revenue-sharing agreement details, and risk factors — the first verified look at ChatGPT unit economics. Anchored against Anthropic's ~$71B compute commitments, TSMC's $265B, and ~$700B in hyperscaler datacenter spend, every AI valuation in the market is currently derived from leaks; the S-1 replaces the comparables set in an afternoon. For TTL the IPO is the single most important market event of Q3 2026: comparables reset, revenue-share structures become public, the consumer-distribution-vs-enterprise-monetization thesis gets answered by public-market multiples within weeks of each other (OpenAI September, Anthropic October roadshow).
Source →openai-ipocomparablessec-filing