8 min read · 13 sources

Nvidia just printed $96.22 billion in a single quarter, more than doubling year-over-year, and told investors to expect $108 billion next. The data-center line alone did $89B. CFO Colette Kress said supply is meeting roughly 70% of demand, which is the only thing stopping the number from being much, much bigger. That bottleneck is the story underneath the headline, and it cascades through everything else this week: data-center bans, memory prices, cloud bills, and the agent economics that Google and Vercel are now racing to make affordable.

Nvidia guided to $108B in Q3 — an 89% YoY jump — but supply is still running at roughly 70% of demand, the company said.

Nvidia's $96B quarter, and the supply wall behind it

Source: apnews.com ↗

Nvidia’s Q2 results beat the $92.27B analyst consensus, with net income of $59.69B ($2.46/share) versus $26.42B ($1.08/share) a year ago. The Q3 outlook of ~$108B implies 89% YoY growth, and the company is guiding to 70% growth for the full fiscal year ending January 2028. That is the kind of number that makes infrastructure planners nervous in a specific way: if Nvidia cannot ship more because TSMC and HBM capacity are full, then the constraint migrates downstream into memory pricing, power contracts, and the willingness of communities to host new data centers. Kress’s “70% of demand” line is the one to remember when reading the rest of the week’s stories about cost controls and moratoria.

CISA gives agencies three days to patch a flaw exploited since March

Source: itnerd.blog ↗

CISA added CVE-2026-21962 to the Known Exploited Vulnerabilities catalog with a remediation deadline of August 27. The CVSS 10.0 bug sits in Oracle HTTP Server and the Oracle WebLogic Server Proxy Plug-in, is remotely exploitable without authentication, and can let attackers read, modify, or delete data. Oracle shipped the patch on January 20, 2026, public exploit code surfaced in March, and CISA’s KEV addition came 216 days after the original fix. Shodan shows roughly 79,000 Oracle HTTP Server instances exposed to the internet. Federal agencies that delayed this one now have a three-day window, plus the forensic-triage requirements of BOD 26-04, to remediate a critical, already-weaponised flaw on systems that should have been patched in Q1.

Salesforce puts the CRM inside Claude

Source: venturebeat.com ↗

Salesforce and Anthropic launched Claudeforce, wiring live CRM data and Salesforce workflows into Claude. The first drop includes 37 sales skills covering pipeline analysis, account research, and deal reviews. The pitch is that an account exec never opens the Salesforce app again. For engineers, the interesting part is what this implies about the agent surface area: vendor-specific skills as protocol-level constructs, identity and authorisation moving into the model layer, and the CRM as a set of tools the LLM can call rather than a UI. It is the first concrete sign that MCP-style integrations are becoming the default delivery vehicle for enterprise software.

MCP stops pretending to be its own transport

Source: pub.towardsai.net ↗

MCP’s 2026-07-28 revision is not a “v2” — it is an explicit pivot to HTTP. The change removes protocol handshakes and persistent sessions, pushes state into requests and server-side data models, and standardises an authorisation profile, tool-risk annotations, and embedded interfaces. The practical consequence for operators is that an MCP server now deploys like any other HTTP workload: behind a load balancer, with normal auth, with normal observability. That is a much bigger deal than the version-number suggests. Teams that held off on MCP because running it in production looked bespoke now have a path that maps onto existing infrastructure. The Claudeforce launch lands the same week, and the two are clearly parts of the same move.

Data-center bans are the new constraint

Source: hpcwire.com ↗

Local opposition to AI buildouts is hardening into actual moratoriums, driven by electricity costs, water use, and grid strain. For an industry that has been treating compute as infinitely expandable at the margin, this is the first serious capacity ceiling that is not about silicon. Combined with Nvidia’s supply comment, it means two of the three levers for growth (wafer allocation, siting) are now binding at the same time. Capacity planners who model GPU availability as the only constraint are about to be wrong.

AWS buys DuckLabs, keeps DuckDB open

Source: aws.amazon.com ↗

Amazon has signed a definitive agreement to acquire DuckLabs, the Amsterdam company behind DuckDB. Founders Hannes Mühleisen and Mark Raasveldt continue leading the project under MIT, and governance sits with the independent DuckDB Foundation. DuckDB’s niche is fast in-process analytics on roughly 1TB or less, with vectorised execution — the same shape of workload that powers a lot of embedded feature engineering. If you currently run DuckDB against S3, expect tighter integration, probably first-party AWS service, and likely managed compute pricing. The licence and foundation structure mean there is no fork pressure yet, but the clock on “AWS-aligned by default” has started.

Vercel ships a sandbox for untrusted agent code

Source: vercel.com ↗

Vercel released the Run SDK (pnpm add run) for executing untrusted JavaScript and type-stripped TypeScript inside QuickJS worker threads. There is no Node.js, no network, and no access to application secrets. Applications expose narrow hostFunctions — regular functions callable as globals inside the sandbox — and calls cross the boundary via serialisation. A run can be paused for auth or human approval and resumed with recorded results. This is what powers “code mode” in Vercel’s AI SDK, and the broader pattern is the one to copy: agents that generate programs to describe API calls and glue logic, executed in a hard sandbox, with the host application keeping custody of every side effect. If you are building agentic systems that touch real systems of record, this is the architecture.

Google Cloud gets serious about agent cost controls

Source: cloud.google.com ↗

Google Cloud shipped a bundle of billing changes aimed at agent workloads. PAYG Gemini Enterprise pricing removes quota limits, Flexible Savings Plans offer 10–20% off token costs with no minimums, and consolidated spend guardrails add hard monthly caps plus runtime cost estimation. Google Antigravity and Android Studio AI usage now roll up into a single Gemini Enterprise bill. For teams running steady agent traffic, the savings plans are the obvious win; for everyone else, the hard caps and cost-estimation tools are the way to keep a runaway loop from lighting the budget on fire.

Memory may eat 68% of cloud capex

Source: theregister.com ↗

A brief Register summary reports that DRAM and NAND could climb to roughly 68% of cloud provider capital expenditure as AI demand tightens the market. The Register’s piece appears truncated, so the underlying model and assumptions are not visible — worth flagging rather than treating as gospel. The directional signal, though, is consistent with the rest of the week: silicon supply is binding, memory supply is binding behind it, and the cost is being passed through into instance pricing and reserved-capacity terms. If you are sizing a multi-year cloud commitment, that 68% number is the kind of thing that should appear in your assumptions.

Databricks' AI SRE investigates 100s of services and 1500+ clusters

Source: databricks.com ↗

Databricks published a writeup of its AI SRE system, an agent that triages incidents across hundreds of microservices and 1,500+ Kubernetes clusters spanning 70+ regions and three clouds. It correlates signals across the stack, executes runbooks, and steers on-call engineers toward root cause. The interesting engineering detail is that the team started with extensive interviews of on-call engineers to map the debugging journey before building any agent. The thesis: the cognitive work of connecting debugging signals is the bottleneck, not the data. For teams running large distributed systems, the pattern — interview the humans first, encode the workflow, then automate — is the part worth stealing.

A strategy piece worth reading even if you don't read strategy pieces

Source: platforms.substack.com ↗

A Substack on “Strategy in the Age of AI” argues that AI changes the rules of competition rather than just improving operational efficiency. The core claim: as AI makes intelligence abundant within individual modules, advantage moves to whoever can connect activities and learn across boundaries. Vertical integration returns, but as ownership of feedback loops rather than physical assets. The warning that lands for engineers: operational excellence can become strategically destructive when the underlying economic value of the work itself is shifting. Read it once and the next reorg memo you see will make more sense.

Gemini Enterprise for financial services lands with auditability front and centre

Source: cloud.google.com ↗

Google Cloud launched Gemini Enterprise for Financial Services, built for capital markets and corporate banking. The package includes 50+ purpose-built financial skills, secure MCP connectors to licensed data sources, a Google-managed Financial Research agent with full explainability, and an open ecosystem with governance underneath. Existing data entitlements and permissions are preserved. The emphasis on verifiable data lineage, explicit confidence scores, and audit trails is the interesting move: it codifies what “production-ready” actually means in a regulated industry, and the same shape is going to leak into adjacent verticals.

RPM, pristine sources, and why your spec file looks like that

Source: bubble.ro ↗

A history of RPM traces the package manager’s design philosophy to “pristine sources”: unmodified upstream tarballs plus separate, reviewable patch files. Erik Troan and Marc Ewing wrote it in 1997, and the choice let distributions rebase onto new upstream versions by re-running builds rather than hand-merging forks. Security backports became auditable as small spec-file diffs. The dependency resolver lived outside RPM for two decades before DNF replaced Yum’s solver with libsolv while preserving the format. If you maintain packages, the philosophy — not the tool — is what explains the shape of the spec files you keep editing.