Executive Summary
Infrastructure took the money again this week, and no flagship model landed. NVIDIA printed $96.2 billion of second-quarter revenue, $89.0 billion of it data centre, and guided the next quarter to $108 billion plus or minus 2 per cent with no China data-centre compute in that number. Alibaba closed an HK$80 billion placing, now cash, split about 60/40 between global compute and hyperscale “Agentic Cloud” data centres. A reported 15–17 per cent rise on early-2027 NVIDIA AI servers is still a leak: it is not in the Q2 release. In Europe, the Dutch data-protection authority fined Uber almost €825 million for firing drivers with no human in the loop. OpenAI published its full write-up of the July Hugging Face breakout.
The tools moved into rooms you already pay for, and a few cheap defaults got more expensive. Slack and Teams can now host a coding agent. Claude can drive Chrome, or its own desktop browser, and it remembers across chat and Cowork. Cursor Auto is no longer a flat rate. OpenAI cut GPT-5.6 Sol’s official list, and the gateway half-price still stacks on top until 18 September. For most businesses the job is still small and concrete: one browser task with approval left on, one honest press of LinkedIn's slop button, and one free Claude Academy lesson run against a real task.
What to Try This Week
Put Claude on one trusted website, with auto-approve off
If you have a paid Claude plan, open Claude in Chrome or Cowork’s built-in browser. Pick one vendor portal you already use (invoices, bookings). Ask it to collect a defined set of records into a spreadsheet. Leave auto-approve off for the first run. Do not use banking, email or anyone else’s data. Measure time saved, errors, and anything that still needed a click.
Press LinkedIn's slop button, then bring the standard to work
LinkedIn now lets members flag a post as AI slop. Give it a go. However, I also encourage you to apply the same test inside your own team: if someone sends you an obviously AI written document, they have clearly not read themselves, send it back unread. Unreviewed AI output moves creates work rather than removing it, and the time saved upstream lands on the poor soul who has to review it. AI as a drafting co-pilot is fine, pages of unread crap isnt (I may feel a bit too strongly about this).
Spend one lunch hour in Claude Academy
Anthropic opened up academy.claude.com this week at no cost, and not all of it is Claude-specific, so parts are relevant to whichever tool you have standardised on. Pick something close to your actual job, take one lesson, then deploy this new super power on a real task (rather than the worked example). Measure whether the second attempt needs less rework than your usual prompt.
Geopolitics, Governance and Big Moves
NVIDIA’s quarter is still compounding, and China is carved out of the guide
NVIDIA’s second quarter of fiscal 2027 (ended 26 July) printed $96.2 billion of revenue, up 18 per cent quarter on quarter and 106 per cent year on year. Data centre was $89.0 billion. Gross margin was 75.0 per cent. The company guides the next quarter to $108.0 billion plus or minus 2 per cent, and says that number includes no China data-centre compute. Jensen Huang said Vera Rubin is in full production. The release does not address Saturday’s report that some large customers were told early-2027 AI servers (Vera Rubin, Grace Blackwell) will cost more than 15 per cent more, later tightened by The Information to about 17 per cent. Treat the hike as unverified. None of this changes what an SME pays for tokens this week.
Source: NVIDIA IR
Alibaba’s HK$80 billion AI placing is now cash
Alibaba priced 710 million new shares at HK$112.70 on 23 August and closed the placing on 26 August. About 60 per cent of net proceeds (HK$47.9 billion) is for global computing infrastructure; about 40 per cent (HK$31.9 billion) for hyperscale AI data centres and an “Agentic Cloud” upgrade of storage, databases and networking. Sold only to non-US persons. The Qwen owner is ring-fencing a balance-sheet raise for capex.
Sources: Alibaba completion notice, pricing release
OpenAI published its full write-up of the Hugging Face breakout
In July, during internal cyber evaluations, OpenAI models with reduced safeguards got around isolation and reached OpenAI research infrastructure and Hugging Face’s systems. On 26 August OpenAI published the full technical report, with CrowdStrike as an external advisor. METR and Redwood Research published a separate alignment investigation the same day. OpenAI calls it a warning shot, and says it is tightening sandboxes, weight access and chain-of-thought monitoring, in part because of this incident and, separately, the capabilities of an upcoming model it names Astra. Customer ChatGPT and API products were not affected. Last week’s Pulse had the training-hold; this is the documented case.
Source: OpenAI
The UK Sovereign AI Fund’s first cheque went to a compute router
London startup Callosum announced a $100 million seed on 20 August, led by Atomico, with Plural, DCVC and the UK Sovereign AI Fund. No valuation disclosed. It says it splits workloads across models and chips, and named a flagship Cerebras partnership. Callosum claims it is the fund’s first investment. Useful as a signal that UK industrial policy is paying for routing and utilisation rather than another foundation model.
AWS and NVIDIA talked about 2 million extra GPUs in 2027–2028
On earnings day the two companies said AWS plans to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs in 2027–2028, on top of an earlier plan for more than 1 million from 2026. Also listed: Vera CPU on AWS, and 100,000 GPUs planned on secure AWS for US federal work. This is a capacity plan, not a booked purchase order, and it does not move this month’s cloud invoice.
Source: NVIDIA
Tools and Releases
Claude Academy is free
Anthropic opened academy.claude.com: courses and badges around Delegation, Description, Discernment and Diligence, with tracks for Claude.ai, Cowork, Code and the platform. Some of it is model-agnostic too (I havent looked but i bet there is nothing on open weight, downloadedable models).
Sources: Claude Academy, Anthropic
Slack and Teams can now host a coding agent
Slack Code (20 August) opens a dedicated channel when you tag Claude, Devin, GitHub Copilot or Vercel: diffs, a live HTML preview, team steering, and an auto-archive audit log. Any Slack plan; you still need the partner agent connected. ChatGPT is listed as coming later. The same week GitHub put Copilot in Slack and Microsoft Teams in public preview. In Slack, @GitHub can plan, triage, investigate in a cloud sandbox and open a pull request. In Teams, a channel, thread or meeting chat can steer one cloud-agent session. Public preview is GitHub Copilot Business and Enterprise, not consumer Pro. Admins must enable the cloud-agent policy. Shared sessions create pull requests as the Copilot app, not you. Do not let it merge.
Sources: Slack Code, Copilot in Slack, Copilot in Teams
Claude memory now spans chat and cloud Cowork
One memory across cloud chat and Cowork (this has its pro’s and cons), with editable Topics. On by default for Free, Pro and Max. Off for Team and Enterprise until an owner turns it on. Local Cowork on your machine does not use it. Sensitive topics stay out unless you switch them on. Team admins should decide, not drift.
Ask Gemini takes over Google Chat, gradually
From 26 August, Ask Gemini becomes Chat’s command line on English Business Standard/Plus, Enterprise Standard/Plus and Google AI Pro for Education accounts. Not Starter. Rollout is Rapid and Scheduled, up to 15 days. The Gemini side panel in Chat is retired and its history does not migrate. Export under “Gemini in Workspace”, not “Google Chat”. Higher limits through 1 October.
Sources: Workspace Updates, Help
Gemini 3.5 Transcribe is in public preview
File and live streaming speech-to-text in the Gemini line, with speaker ID (up to three speakers) and word timestamps, in more than 85 languages. On Gemini API, AI Studio, and Vercel AI Gateway the same day. Useful for internal meeting notes. Keep client audio out until legal or IT says yes.
ChatGPT Work can wait for Gmail, Slack or a GitHub pull request
Plus and Pro Work seats can fire a scheduled task when a new Gmail arrives, a Slack channel is posted, or a GitHub pull request moves. Keep confirmation on. Separately, ChatGPT Business Premium seats are live: $125 per user per month, or $100 on annual billing, against Standard at $25 / $20. Premium is 5× Standard usage and drops the five-hour cap. Mix seats. Do not upgrade the whole firm because one person hits the wall.
Sources: ChatGPT release notes, Premium seats
DeepSeek’s cheap vision model is live, and experimental
deepseek-v4-flash-vision-exp launched 21 August: screenshots, charts and documents at V4-Flash token prices. Live on DeepSeek’s API, Vercel Gateway and OpenRouter. Images cap at 384 tokens after resize. China-origin: data leaves the EEA unless you already have a residency setup. Vendor benches versus Opus 4.8 are theirs. One real screenshot if you already have a key. Not a production default.
Sources: DeepSeek changelog, Vercel
Claude can drive Chrome, or its own desktop browser
Claude in Chrome is generally available on all paid plans (26 August). It can auto-approve actions it judges safe; switch that off if you want ask-first. Chrome only, not other Chromium, not mobile. Enterprise admins can limit it to approved domains. Prompt injection is acknowledged, not gone. Separately, Cowork gets a built-in browser inside Claude Desktop: no extension, it does not see your tabs or passwords unless you import cookies site by site. Rolling this week to Pro, Max and Team. Enterprise if the owner enables it. The desktop app has to stay open.
Sources: Claude in Chrome, Cowork browser
GPT-5.6 Sol’s official list is now $4 / $20, and the gateway cut still stacks
OpenAI cut short-context Sol to $4 per million input tokens and $20 per million output tokens (was $5 / $30), promotional through at least 21 November. Cached input is $0.40. Long context over 272K is 2× input and 1.5× output for the full request. AWS Bedrock matches. Vercel Gateway’s 50 per cent off still applies to the new list through 18 September, so default tier is $2 / $10 if you are not on bring-your-own-key. OpenRouter still shows the same stack. Existing requests reprice automatically. Do not bake the promo into next quarter’s budget. Azure-routed OpenRouter may still show the old list.
Sources: OpenAI pricing, Vercel, Bedrock
Cursor Auto’s new billing is now on the pricing page
cursor.com/docs/models-and-pricing now says all Auto modes bill at the list price of the model each request is routed to. Third-party routed Auto on Teams and Enterprise also picks up a Cursor Token Rate of $0.25 per million tokens. Legacy Enterprise Auto stays flat until 7 September. The public changelog was still 19 August at the end of this window. Most routed turns will cost more than the old flat rate.
Source: Cursor pricing
ChatGPT ads in 31 European markets are still not confirmed live
Last week’s Pulse said ads would land on 24 August for Free and Go. At the end of this window, Ads Manager still listed all 31 EEA and Swiss rows as Coming Soon. Plus, Pro, Business, Enterprise and Education stay ad-free. The UK has had ads since June. Turning personalisation off does not remove ads. Treat Europe as a status check until the rows flip.
Sources: OpenAI, Ads Manager availability
Quick Hits
Grok 4.6 is on Microsoft Foundry (preview) as well as Bedrock and Google’s Enterprise Agent Platform. Same model, another procurement door. xAI
Z.ai shipped GLM-5.3-Flash weights under MIT (320B total, 18B active). The full GLM-5.3 checkpoint is still not out. Hugging Face
Mechanical Turk closes 30 September. HIT submit ends that day; approve/reject until 30 October. If you still send labelling through it, start the move. Amazon
LinkedIn’s slop button has been used by more than a million members. LinkedIn says posts it classifies as slop get about 40 per cent fewer views. That 40 per cent is theirs. The Verge
Apple’s M6 Mac mini starts at $899 in the US, ships 22 September. Confirm the UK store and sterling price before anyone buys. Apple
Thomson Reuters launched an in-house model for CoCounsel Legal tabular work. UK general availability of the next-gen legal stack is still later this year. Thomson Reuters