The Big Picture
Wall Street didn’t just underwrite AI compute this week — it got underwritten back. When Nvidia’s own five-year CDS nearly doubled in three weeks over circular-financing concerns, the response overnight was Nvidia agreeing to guarantee up to 25% of residual chip value on projects the $500B consortium finances, using what its own 10-Q calls a credit derivative. That is: Nvidia will eat losses on unused chips to keep the buyers safe. This turns the world’s most valuable company into a lender-of-last-resort on its own product, which is either a genius durability play or the point at which “picks and shovels” starts looking like everyone else’s balance-sheet risk. Meanwhile Made by Google tonight tests whether Gemini on a phone can matter as much as Gemini in a datacenter; IBM and Together AI’s $240M inference cluster deal bets enterprise buyers want to bypass frontier-lab pricing entirely; and Cisco tonight tests whether the $9B AI networking orders it teased last quarter actually hold up. The week’s story is that every layer of the AI stack is being restructured under duress at the same time.
Midday refresh: Cisco Q4 answered the morning’s question loudly — $17.3B revenue (+18% YoY, ahead of $16.83B consensus), $4B in AI orders in Q4 pushing FY26 to $9.3B above the raised guide, and a fresh $7.5B FY27 AI-revenue guide that nearly doubles this year’s $4B delivered. That’s the specific forward number the tape needed to see; every AI-earnings print left in 2026 has to clear a higher bar now. Made by Google is running live: Pixel 11 at $899 (256GB base), Tensor G6 on TSMC 2nm, a Santafe TPU pitched at 50% more compute running Gemini Nano 3.5x faster with 3.5x less energy than G5. The tell buried inside the launch: Pixel Watch 5 ships August 20 without its marketed Gemini Intelligence feature — the NPU silicon isn’t ready. Google is selling the on-device-AI pitch on hardware that hasn’t entirely caught up to it.
Live: Made by Google 2026 Special Report →
Made by Google keynote live — Pixel 11 at $899, Tensor G6 on 2nm, Watch 5 ships without Gemini
Today Updated 2:05 PM Google, TechCrunch, Android Authority, 9to5Google, Engadget
Made by Google 2026 kicked off Wednesday August 12 at 3 PM PT in New York. Pricing landed: Pixel 11 starts at $899 (256GB base, 128GB tier dropped), Pixel 11 Pro at $1,099, Pixel Watch 5 at $399 for 41mm / $429 for 45mm (Stephen Curry edition $579), Pixel Tag at $29 individual / $99 four-pack. Tensor G6 confirmed as the first smartphone silicon on TSMC 2nm: seven-core CPU (1 Cortex-C1 Ultra at 4.11 GHz, 4 C1 Pro perf, 2 efficiency), dedicated “Santafe” TPU with 50% more compute running Gemini Nano 3.5x faster with 3.5x less energy vs the G5, “Metis” ISP, Titan M3 security chip. New Gemini surfaces: Rambler voice input (handles run-on speech and filler), Live Transcribe with American Sign Language translation via camera, Circle to Search direct from Camera, Pixel Buds voice-lookup for Pixel Tag. Pre-orders open, general availability August 20.
The tell buried inside the launch: Pixel Watch 5 ships August 20 without its marketed Gemini Intelligence feature — the required NPU silicon isn’t ready, and the feature arrives in a separate update later in 2026. The Gadgeteer’s morning framing that Tensor G6 is subordinate to the on-device AI story landed on the phone side (3.5x/3.5x Gemini Nano is what actually makes local models feel interactive). But Google spent the keynote insisting the hardware-software gap doesn’t exist while its flagship watch demonstrates that on some surfaces it very much does — selling the on-device-AI future on silicon that hasn’t entirely caught up to it.
Nvidia guarantees up to 25% chip residual value to unlock $500B AI financing consortium
Today Bloomberg, Yahoo Finance, Taipei Times, The Edge, The Decoder
Bloomberg reported Wednesday August 12 the mechanical detail Monday’s $500 billion Wall Street financing consortium announcement had left implicit: Nvidia is guaranteeing up to 25% of the residual value of its own chips installed in financed projects, classified as a credit derivative in the company’s own SEC 10-Q. If resale or reuse value at end-of-term falls below expectations, Nvidia covers part of the gap on a project-by-project basis after other recovery steps — new leasing customers, chip resale — are exhausted. Structural context: Nvidia’s five-year CDS nearly doubled in three weeks running up to the announcement as credit markets priced in circular-financing risk from the OpenAI-Oracle vendor-loan loop. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR will independently judge individual deals and set their own participation, with Nvidia’s guarantee kicking in on some but not all transactions.
Bloomberg’s point: the RVM is what makes AI compute an investable asset class in practice, not just in press release. The 25% residual guarantee is the specific mechanism that turns a GPU cluster into collateral rated pension funds can hold. Every subsequent Wall Street pitch for AI-infra financing now has to answer whether the chipmaker guarantees the collateral or not — and Nvidia has effectively agreed to eat losses on unused chips to keep buyers safe. This is either a genius durability play or the point at which “picks and shovels” starts inheriting everyone else’s balance-sheet risk.
Cisco Q4 crushes it — $17.3B revenue, $9.3B FY26 AI orders, $7.5B FY27 AI revenue guide
Today Updated 2:05 PM Cisco IR, CNBC, Yahoo Finance, TradingKey
Cisco reported fiscal Q4 2026 after the bell Wednesday August 12: revenue $17.3 billion (+18% YoY, above $16.83B consensus), GAAP EPS $0.97, non-GAAP EPS $1.22 (above $1.17 consensus). $4 billion in AI infrastructure orders in Q4 alone, bringing FY26 total to $9.3 billion — above the raised $9B annual guide. Networking product orders +40% YoY in Q4 (Q3 was +50%, so growth decelerating but still strong); total Q4 product orders +35% YoY, +25% ex-hyperscalers. Chuck Robbins framed FY26 as “another record year.” Fresh number the tape got: FY27 AI infrastructure revenue guide of $7.5 billion — nearly double the ~$4 billion Cisco actually delivered in FY26 — on backlog already booked.
Cisco’s point: the $7.5B FY27 AI revenue guide isn’t “pipeline,” it’s already been ordered. That closes the argument TradingKey framed this morning — the hyperscaler-fabric TAM isn’t capping, it’s still expanding by double from delivered revenue. Combined with AWS +37%, Palantir US-commercial +149%, and Cloudflare +36% this cycle, CSCO tonight raises the AI-earnings-tape bar Nvidia has to clear August 26. The morning’s cynical case (that a held-flat guide would signal a capped TAM) doesn’t apply — the guide got raised, not held.
IBM and Together AI sign $240M multi-year deal for Nvidia HGX B300 open-source inference cluster
Yesterday IBM newsroom, Reuters via Yahoo, Quartz, Open Source For You, BNN Bloomberg
IBM and Together AI announced Tuesday August 11 a multi-year $240 million agreement to build a large-scale AI inference cluster on IBM Cloud using Nvidia HGX B300 systems — Blackwell-based accelerators with Nvidia Spectrum-X Ethernet networking. IBM described it as the first dedicated large-scale cluster built for inference on IBM Cloud with those systems; availability targeted Q1 2027. The cluster is purpose-built for serving open-source AI models at scale to enterprises hedging against per-token frontier-lab pricing. Together AI, which raised an $800M Series C at $8.3B valuation earlier this cycle, is the workload partner. Framing on Reuters wire cited “concerns about cybersecurity incidents involving models from Anthropic, OpenAI and Meta” as enterprise motivation for open-source paths.
Reuters’ framing: first dedicated large-scale inference cluster on IBM Cloud built specifically for open-source AI models. Direct enterprise pitch is to bypass Anthropic and OpenAI per-token pricing entirely — Together’s open-model catalog on Nvidia B300s becomes the alternative stack. Lands 48 hours after Monday’s NVIDIA Nemotron 3.5 Lightning open weights and General Catalyst’s $1.1B check into River AI: open-source enterprise AI now has three concrete deployment paths (Nemotron self-hosted on Nvidia, River AI fine-tuning-as-a-service, IBM Cloud Together inference) all landing in one week. Pre-IPO pressure on Anthropic and OpenAI just got a third pricing lever added to it.
Tencent WorkBuddy floods to top of China workplace AI — Bloomberg calls it turnaround moment
Yesterday Bloomberg, TechNode, Tencent Hy3 release, Winbuzzer
Bloomberg reported Tuesday August 11 that a flood of users has made Tencent’s WorkBuddy the most popular AI workplace assistant in China, giving the internet giant its first credible catch-up move after months of trailing peers in the domestic AI race. WorkBuddy runs on Tencent’s Hy3 model (Apache 2.0, 295B MoE / 21B active, 256K context, released globally July 6) and executes natural-language office tasks — reading documents, analyzing spreadsheets, drafting reports, generating posters, orchestrating multi-step MCP workflows through 20+ skill packages. Hy3 access through WorkBuddy is free worldwide through August 31 PT. Global launch was in May 2026; the traction Bloomberg reported is the domestic-China acceleration since June.
Bloomberg’s framing: WorkBuddy adoption is the first concrete data point in Tencent’s post-rout AI turnaround narrative — the stock had been punished for months on questions about whether it could compete in agentic AI at all. Specific tell: Tencent is running the same August-31 free-tier acquisition play OpenAI ran with ChatGPT Business Premium credits Monday. Two labs in different jurisdictions both racing to convert curiosity into locked-in seats before Q4 pricing turns on. The retention numbers when Hy3 stops being free in September will decide whether WorkBuddy is a turnaround or a promo bump.
OpenAI ships ChatGPT desktop app for Linux public preview — ChatGPT, Work, and Codex bundled
Yesterday TechCrunch, releasebot, iGeeksBlog, OpenAI newsroom
OpenAI released ChatGPT desktop app for Linux in public preview Tuesday August 11, bundling ChatGPT, ChatGPT Work, and Codex into a single native binary shipped as DEB and RPM packages — the first Linux client OpenAI has ever published. Feature parity with the macOS and Windows desktop apps is the launch claim: keyboard-triggered inline queries, file drag-and-drop, multi-window conversation state, Codex sidebar for code sessions. Enterprise SSO parity with the macOS build. Skips Snap and Flatpak entirely; the target is enterprise Linux workstations, not consumer distros. Public preview is free to download for any paid ChatGPT tier user.
TechCrunch’s framing: first native Linux client OpenAI has ever shipped — the specific tell that ChatGPT is now hardening its developer-facing surface. Bundling three products into one desktop app is the same play Anthropic ran with Claude Code plus Cowork — the frontier lab that owns the developer inbox on the workstation owns the seat. DEB and RPM without Snap/Flatpak is the tell that this is aimed at enterprise Linux fleets running RHEL, Ubuntu, and Debian in production — not hobbyist installs.
Anthropic extends Claude Code 50% weekly usage boost through August 19 for Pro, Max, Team
Yesterday Anthropic support docs, TestingCatalog, AI Builder Club, DevBytes
Anthropic extended the temporary 50% weekly usage boost for Claude Code through 11:59 PM PT on August 19, 2026, per an updated support document Tuesday August 11 — the third extension of the same promo since it launched May 13. Applies to Pro, Max, Team, and legacy seat-based Enterprise users automatically across CLI, IDE extensions, desktop, and web surfaces. Free plans and consumption-based Enterprise seats are excluded. Five-hour usage cap stays unchanged; only the weekly ceiling moves. Lands two days before Auto Mode becomes the default permission mode in Claude Code for every non-Enterprise paid user on August 14.
TestingCatalog’s read: third extension since May and Anthropic still hasn’t turned it off. Specific tell — five-hour burst cap unchanged, only the weekly ceiling moves. That means the constraint Anthropic is subsidizing is total spend per week, not burst rate: the workload it’s protecting is long-running background agents that would otherwise hit the weekly ceiling by Thursday. Auto-mode-as-default lands August 14, the same week the boost expires. Sequencing looks intentional: give users the ceiling headroom to actually feel Auto Mode’s throughput advantage before flipping the default.