Thursday, August 27, 2026

Nvidia's Q2 print landed, and the number that matters isn't the beat — it's that Nvidia gave a year-ahead forecast for the first time in its history: 70% revenue growth in FY28 against a 44% consensus. In the same 24 hours AWS committed to 2 million more GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗, and Anthropic signed a $45B six-year lease. Underneath the headlines there's one thing your team can actually act on today: a critical, still-unpatched prompt-injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ in Context7, one of the most widely installed MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ documentation servers.

TLDR

Nvidia reported $96.2B in Q2 revenue (+106% YoY), guided Q3 to $108B, and — for the first time ever — gave a year-ahead forecast: ~70% revenue growth in FY28 vs. the 44% analysts expected. Supply commitments more than doubled quarter over quarter, $119B → $279B (Aug 26) Infra Trending
AWS and Nvidia committed to 2 million additional GPUs in 2027–28, Vera CPUs on AWS, and 100,000 GPUs for US government AI factories — and, buried in the same release, GPU-accelerated vector indexing on Amazon OpenSearch at 9× faster index builds for a quarter of the cost (Aug 26) Infra Data
Anthropic signed a ~$45B, six-year compute lease with Nscale, running on Nvidia Vera Rubin out of a West Virginia data center, live from late 2027 (Aug 26) Infra
Alibaba open-sourced Qwen3.8-Flash-Next — a ~125B MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ released as a preview of the Qwen 4 architecture, weights on Hugging Face and ModelScope (Aug 26) Open Source
CVE-2026-75130: a critical prompt-injection flaw in Context7's "Custom AI Instructions" plants attacker-controlled text into any connected coding agent — and no public fix is documented (published Aug 18) Security Agents

Action Items

Things your engineering team should look at or do this week.

Inventory which MCP servers your coding agents are wired to, and check whether Context7 is one of them. CVE-2026-75130 affects Context7 2.1.2 and earlier, scores 9.0 critical under CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 3.1, and had no documented public fix as of publication. The attack path is mundane, which is what makes it dangerous: your agent makes a routine library-documentation request, and unsanitized "Custom AI Instructions" ride back into its context. Documented impact is credential exfiltration from .env files and destructive file deletion. Treat every MCP server as an untrusted content source that reaches straight into your agent's instruction channel — this is the third distinct "the thing that configures your agent is the attack surface" story in ten days.
Migrate off gemini-robotics-er-1.6-preview before Monday, Aug 31. Google shuts it down at the end of the month; the replacement is gemini-robotics-er-2-preview (public preview since Jul 30), with a streaming sibling for real-time text. Small blast radius for most teams, but a hard stop if anything in your spatial-reasoning or robotics stack still points at 1.6.
If you run billion-scale vector search, benchmark GPU-accelerated indexing on OpenSearch before your next capacity conversation. AWS is claiming up to 9× faster vector index building at roughly a quarter of the cost, on both managed clusters and Serverless, plus 3.7× faster data processing and 30% better price-performance for cuDF on EMR with G7 instances. Index-build time is the bottleneck nobody budgets for in RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ pipelines, and these are vendor numbers on a well-defined workload — cheap to verify against your own corpus.
Re-check your Anthropic API failover posture. Claude.ai logged another elevated-error incident yesterday (13:52–14:45 UTC), following Aug 5, 12, 13, 16, 18, 20 and 24. Whatever your monthly reliability assumption for a single-provider dependency was in July, August has given you enough data to revise it.

Models & Tools

Alibaba open-sources Qwen3.8-Flash-Next — a Qwen 4 architecture preview, shipped as open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ (Aug 26) Open Source — A Mixture-of-Experts model at roughly 125B parameters, positioned as the cheap, fast tier rather than a flagship, with weights published to Hugging Face and ModelScope — no proprietary API required. Alibaba claims performance competitive with Anthropic's Opus 4.6 and DeepSeek V4-Flash, and says the architecture "significantly reduces training and inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ cost." The strategically interesting part is the labeling, not the benchmarks: Alibaba is explicitly calling this a preview of the Qwen 4 architecture, which means the next generation's efficiency story is being validated in public, under open weights, before the flagship exists. Treat the head-to-head comparisons as vendor claims until third parties reproduce them. (Bloomberg | TechNode)
AWS × Nvidia: 2M more GPUs, Vera CPUs on AWS — and a real data-layer speedup hiding in the press release (Aug 26) Infra Data — The headline commitment is 2 million additional Blackwell Ultra, Rubin and Rubin Ultra GPUs across AWS in 2027–28, on top of the 1M+ announced at GTC 2026 — which AWS says demand already outran. Also in scope: Nvidia Vera CPU infrastructure on AWS for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workloads that need heavy CPU alongside accelerators; NVLink Fusion with custom NVHBM memory extended to Trainium via Annapurna Labs; 100,000 GPUs on secure AWS infrastructure for US federal and national-security workloads at IL6 and above; Nemotron open models staying on Bedrock and SageMaker; and Amazon Robotics adopting Nvidia's Jetson/Omniverse/Isaac physical-AI stack. The line worth your attention as an engineer rather than a market-watcher is the data layer: GPU-accelerated vector indexing on Amazon OpenSearch, up to 9× faster index construction at ~25% of the cost, and cuDF on Amazon EMR with G7 instances at up to 3.7× faster processing and 30% better price-performance. Those are the two costs that actually show up in a RAG bill. (NVIDIA Newsroom | About Amazon | TechCrunch)
Nvidia Jetson Orin Nano 2: 2× the inference in the same form factor, or the same performance at 40% less power (Aug 25) Robotics/IoT Infra — The entry-level edge module gets 78 TOPS, 8GB LPDDR5x at 120GB/s, and an 8-core Arm CPU, keeping the Ampere architecture and the existing footprint. Nvidia's two framings are the useful ones: double the inference performance at the same size, or the previous Super's performance at 40% less power in a 15W envelope. Two caveats that matter for planning: no price has been announced (the $249 Orin Nano Super remains the entry point), and module and dev kit availability is H1 2027 — so this is a roadmap datapoint for anyone doing on-device vision or drone work, not something to design against this quarter. (NVIDIA Newsroom | Unite.AI)
Visual Studio 18.9 adds per-request thinking-effort control to Copilot (Aug 25) — Alongside next-edit-suggestion improvements, the release lets you pick low / medium / high reasoning effort per supported model: low for completions and straightforward questions at minimal tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ cost, high for gnarly algorithms, architecture calls, and hard debugging. Small feature, real budget lever. Most teams currently pay flagship reasoning prices on requests that never needed reasoning, and an in-IDE dial is a more honest control than a model-routing policy nobody follows. (Microsoft Learn — VS 2026 release notes)

APIs & Pricing

Google: gemini-robotics-er-1.6-preview shuts down Monday, Aug 31 Pricing Robotics/IoT — Released Apr 14, superseded Jul 30, and now retired. The migration target is gemini-robotics-er-2-preview for spatial reasoning and multi-robot coordination, with gemini-robotics-er-2-streaming-preview for real-time streaming. Narrow audience, hard deadline — four days out. (Gemini API changelog)

Security

CVE-2026-75130 — a 9.0 prompt injection in Context7, one of the most widely installed MCP documentation servers, with no documented fix (published Aug 18) Security Agents — The flaw sits in Context7's "Custom AI Instructions" feature, which places unsanitized text into any connected coding agent's context when the agent makes an ordinary library-documentation request. Classified CWE-1427 (improper neutralization of input used for LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ prompting), credited to Eli Ainhorn of Noma Security. Documented impact: credential exfiltration from environment files and destructive file deletion. Affected range is 2.1.2 and earlier, and no public fix was documented at publication. Note the scoring split, because it will come up in your triage meeting: 9.0 critical under CVSS 3.1, 6.4 medium under CVSS 4.0 — the same bug, opposite verdicts, which is a good argument for judging agent-context bugs by blast radius rather than by score. It's more than a week old, but it remains unfixed and Context7 is installed in a very large number of developer agent configurations. (The Hacker Wire | NetFoundry CVE/KEV tracker)

Industry & Business

Nvidia's Q2: a beat, and then a first-ever year-ahead forecast that reframes the whole AI-capex debate (Aug 26) Infra Trending — Revenue $96.2B, up 18% sequentially and 106% year over year, against a ~$92.2B consensus. Non-GAAP EPS $2.22 vs. $2.06–2.09 expected. Data center $89B (+117%). Gross margin held at 75% for a second straight quarter, under real pressure from memory and wafer costs. Q3 guidance: $108B ±2%, against ~$104.2B expected. The genuinely new thing is the forward guide. CFO Colette Kress told analysts to expect ~70% revenue growth in FY28 — the company's first year-ahead forecast ever, against a 44% consensus — and Huang said supply, particularly memory, is the constraint on going higher. Supply commitments more than doubled from $119B to $279B, mostly memory procurement, and Kress put top-five-hyperscaler capex at $800B in 2026 rising to $1.3T in 2027. Why a team lead should care about a chip company's guide: that forecast is a claim about token prices two years out. If it's right, the cost curve you're planning inference budgets against keeps falling; if the memory constraint bites, the recent run of price cuts is a floor rather than a trend. (Fortune | CNBC | NVIDIA IR)
Anthropic leases ~$45B of compute from Nscale — a company founded in 2024 (Aug 26) Infra — Per Bloomberg and TechCrunch sources: a six-year agreement, compute delivered on Nvidia's Vera Rubin platform from Nscale's West Virginia flagship data center, powering Anthropic services from late 2027. It extends a remarkable eight-month run — $10B with Volta (Aug 4), $5B with AMD (July), a SpaceX deal reportedly worth $1.25B/month (May), 5 additional gigawatts from Amazon and an expanded Google/Broadcom TPUTensor Processing UnitGoogle's custom chip built to speed up machine learning, especially large matrix operations.Google Cloud ↗ arrangement (April). The structural observation worth carrying: the counterparties are increasingly two-year-old infrastructure startups rather than hyperscalers, which spreads Anthropic's supply risk across balance sheets that have never been tested through a downturn. If you depend on Claude in production, that's a supply-chain fact about your stack, not just a finance headline. (TechCrunch | Bloomberg)
Emerald AI raises $150M Series A at $1.05B to make data centers flex against the grid (Aug 25) Infra — Oversubscribed, co-led by Energize Capital and DCVC, taking total funding past $220M, with 12 Fortune Global 500 companies now on the cap table. The software lets data centers cut power draw when utilities are strained and soak up excess capacity when it's available, instead of the binary "run or shut down." Emerald claims flexible-load techniques could unlock 100+ GW of untapped capacity on the existing US grid, and says it's now deployed commercially at multi-megawatt, full-data-center scale after five commercial demonstrations. Read it next to Nvidia's $1.3T hyperscaler capex number: the binding constraint on AI buildout is drifting from chips to interconnect queues and megawatts, and this is the software category that arbitrages the gap. (Business Wire | Emerald AI)
OpenAI expands free ChatGPT for Teachers to 55 more school systems — and solves the contract problem with one agreement across 16 states (Aug 26) Customer AI Policy — The expansion adds 100,000+ educators and staff across 20 states, bringing the program to 100+ K-12 organizations in 30 states and 300,000+ educators. Access stays free for verified US K-12 educators through June 2028 and remains restricted to administrators, faculty, and educators. The procurement mechanism is the part with generalizable signal: OpenAI signed a data-privacy agreement spanning 16 states, so districts validate the tool against their own student-data rules through one shared process instead of negotiating individually. Multi-jurisdiction data agreements as a distribution strategy is a pattern that will show up well beyond education. (9to5Mac | OpenAI)

AI in Production

Claude.ai logs another elevated-error incident — the eighth flagged day this month (Aug 26) Production — Anthropic opened an investigation at 13:52 UTC, with errors back to baseline by 14:45 UTC (07:45 PT) and the incident closed shortly after — a shorter event than Monday's 2h46m multi-model outage, and no post-mortem published. Counting the publicly logged days, August now includes Aug 5, 12, 13, 16, 18, 20, 24 and 26. The takeaway isn't any single incident; it's the base rate. Roughly one flagged day in three is a different reliability assumption than most teams encoded when they made Claude a hard dependency, and it's worth putting a number on your own error budget for provider-side failures rather than inheriting one. (Anthropic Status)

Yesterday was the compute-money day.