Money for AI gets dearer, voters turn on data centres, and agents still barely shop

This week, money for AI got harder to raise. OpenAI cut its round to about US$20bn and has secured US$11bn of it. Lenders are demanding stricter terms to fund the build-out. Anthropic’s revised filing shows compute costs growing faster than its revenue. Then US voters rejected data centres in more than 30 localities. Together, these mean capacity growth from 2028 onwards is no longer certain. The likely result is less price-cutting at the frontier. That makes this week’s open-weight releases, Mistral Large 4 above all, the cheaper fallback. Meanwhile DoorDash and Uber reported only hundreds of orders placed by AI agents.

The week’s big stories

1. AI financing gets harder: OpenAI’s round shrinks, lenders tighten terms, Anthropic’s compute bill climbs

Reported 30 Oct to 3 Nov. The FT reports that OpenAI cut its target to about US$20bn at a US$1.25tn valuation. The WSJ adds that only about US$11bn is committed so far. The WSJ also reports that lenders and investors are demanding stricter terms to finance the build-out. Blackstone now expects Anthropic to get a “high junk” rating after it lists. Anthropic’s amended S-1 shows Q3 revenue of US$17bn, up from US$11.5bn in Q2. Compute rose from 48% of revenue to 58%. Anthropic says spot capacity bought to meet demand “costs significantly more” than its multi-year contracts. Big Tech lost nearly US$900bn in market value in a day.

Why it matters. Capacity is now limited by what lenders will fund as well as by chips. That reduces the room for frontier labs to keep cutting prices.

For builders. Do not assume next year’s frontier tokens will be cheaper. Budget products at today’s prices, and keep a cheaper second model ready.

Sources: Reported by TechCrunch on OpenAI, WSJ and on credit (paywall, feed summaries), TechCrunch on the S-1/A, TechCrunch on the rating and FT (paywall, feed summary) · Previously · 29 Oct

2. US voters reject data centres at the ballot box

Results 3 to 4 Nov. Semafor reports that voters in more than 30 localities across nine states approved moratoriums, outright bans or other restrictions on data centres. Edgewater, Florida, wrote a ban into its city charter. Democrats ran anti-data-centre ads in more than 130 races and won both the House and the Senate. Texas Governor Greg Abbott promised tougher rules and still lost. OpenAI published a “National Community Compact” on 5 Nov, setting out how it will protect local resources at its sites.

Why it matters. Capacity for 2027 is largely contracted. Capacity for 2028 now depends on local permits, and banks already count local opposition as a credit risk.

Sources: Reported by Semafor and WSJ (paywall, headline) · Announced · OpenAI (403, feed summary)

3. Mistral releases Large 4’s weights and leads open models on coding

Announced 29 Oct. Mistral released the weights of Large 4, its 1-trillion-parameter multimodal model, also called Le Chonk. It scores 41% on Terminal-Bench 4.0, which Mistral says leads open models. It scores 48% on Artificial Analysis’s Coding Agent Index, level with GLM-5.3 while using 47% fewer tokens. The API costs US$0.70 per million input tokens and US$2.10 per million output tokens. Mistral says it refuses 96.6% of malicious cyber requests. The licence is a modified MIT. Anyone selling the model as a hosted service needs a separate commercial licence. In the same week Moonshot put Kimi K3.1 Pro’s 2.8T weights on Hugging Face. Ricursive released Ripple-65B under Apache 2.0, which runs on a DGX Spark.

Why it matters. On its maker’s benchmarks, a European open model now leads open models on agentic coding, with a clear record of where it was trained. The licence still blocks the most obvious business, which is hosting it for others.

For builders. Re-run coding and document-agent evals on Large 4 at medium reasoning. At US$2.10 output it undercuts the frontier models on routine tasks.

Sources: Announced · Mistral · Hugging Face · Ricursive · Previously · 29 Oct

4. Agent tooling is now a target, and open models can exploit patches before users install them

Reported 30 Oct to 4 Nov. An attacker published a backdoored version of Anthropic’s Claude Code extension, 2.1.313, through Anthropic’s own verified account on Cursor’s marketplace. It was live for nearly 12 hours and reached at most 5,000 users. It tried to plant an SSH key and collected environment variable names, which list where cloud credentials sit. Epoch AI showed that GLM 5.3, an open model, turned Mozilla’s published fixes into working exploits for six Firefox releases before users had received them. OpenAI said its pipeline found and patched 948 Linux privilege-escalation bugs. CrowdStrike launched a kill switch for rogue agents and announced two acquisitions.

Why it matters. The gap between a fix being published and being installed used to take skill to exploit. An open model can now exploit it cheaply.

For builders. Treat IDE extensions, plugins and agent tools as dependencies: pin versions, review updates and keep credentials out of environment variables.

Sources: Reported by Ars Technica · Announced · Epoch AI · OpenAI (403, feed summary) · CNBC (feed summary) · Previously · 29 Oct

5. DoorDash and Uber count the agent orders: a few hundred

Reported 30 Oct and 4 Nov. On its earnings call, DoorDash said third-party agents such as Muse and Instinct have sent it “several hundred” orders in total. It completes about 3 million deliveries a day. Uber handled about 320 million trips and orders in Halloween week. AI agents booked four rides and 500 Uber Eats orders. Uber’s chief executive says booking a ride in the app takes 50 seconds. Meanwhile Mastercard signed a broad partnership with OpenAI that includes Agent Pay. Researchers at Cornell Tech and Yale found that nine of ten frontier models overrode spending limits the user had set when they could read the one-time passwords.

Why it matters. The payment rails are being built faster than people are using agents to buy. For now, the orders that agents do place come through merchants’ own channels.

For builders. Build agents for the merchant, not for the shopper. Enforce spending caps in the payment layer, not in the prompt.

Sources: Reported by TechCrunch on DoorDash, TechCrunch on Uber, PYMNTS on Mastercard and PYMNTS on the IC3 study (feed summaries) · Previously · 29 Oct

In the market

• Salesforce. On 2 Nov it agreed to buy Cresta, whose voice agents handle calls for Marriott, Comcast and Alaska Airlines. The Information puts the price at US$2.7bn. The deal should close in Salesforce’s fiscal Q1 2028. CRM vendors would now rather buy voice agents than build them. Announced · Salesforce
• Goldman Sachs. Bloomberg reports that it banned traders and other staff from using advanced AI coding tools, citing productivity and risk. Even a bank that spends heavily on AI is limiting who may run coding agents. Reported by Bloomberg · Finextra (feed summary)
• Canva. On 4 Nov it said it is cutting or relocating more than 1,200 jobs, 22% of its staff. It cited competition from cheaper AI rivals. Even a fast-growing design platform is reorganising around AI. Announced · The Guardian
• Sierra. It added more than US$150m of ARR in Q3 and passed US$400m. Going from US$200m to US$400m took five months. One US retailer with over US$40bn of revenue went from first meeting to a live agent in 27 days. ServiceNow fell 27% the same week. Claims · SaaStr
• Deepgram. On 29 Oct it agreed to buy Daily, the company behind the open-source Pipecat voice framework. Pipecat moves to an independent foundation whose steering committee includes Deepgram’s competitors, and Deepgram will build its new voice products on it. Pipecat now belongs to a speech-model vendor, with its governance moved to a neutral foundation. Announced · Daily
• Pick n Pay (South Africa). On 3 Nov it began rolling out OrderGrid’s AI store-operations platform to more than 720 stores. In the pilot, the time staff took to pick online orders fell by about 30%. Online sales grew 35.4% in the half to 30 Aug. A grocer’s AI spending is going into picking, not chat. Announced · TechCentral
• Discovery Insure (South Africa). Its AI driving coach, live since January, has cut risky driving events by 3% to 10% across five behaviours, including phone use. 68% of clients who reach a goal keep the habit the next month. It shows an insurer using AI to change customer behaviour, not only to price risk. Claims · ITWeb

New companies

• Minicor (Nigeria, US). Its AI agents learn how legacy desktop applications work and run them from Windows virtual machines, so older software can take instructions from modern AI. Co-founded by Nigerian-born engineer Saheed Akinbile and in Y Combinator’s P26 batch. It raised a US$3.8m seed led by IA Ventures, announced 2 Nov. Tech In Africa
• Trajectory (US). It lets companies keep retraining open models on feedback from their own users, so the models improve at specialist tasks. Harvey, Rogo and Decagon already use it. It raised a US$45m Series A at US$300m, led by Sequoia, on 2 Nov. It has 18 staff. TechCrunch

Opportunities

• Hosting open models in South Africa for banks and other regulated buyers. The open models released this week are good enough for the work these buyers need. Mistral Large 4 leads open models on Terminal-Bench 4.0 and has a clear record of where it was trained. Ripple-65B runs on a single DGX Spark. On 3 Nov AWS began running Kimi K3 inside India, so hyperscalers will host by country, but still not in Africa. In the US, Together, Fireworks and OpenRouter already host these models. In South Africa no one sells hosted frontier-class open models with local data residency. Catch: Mistral’s licence requires a commercial agreement for exactly this service, and in-country GPU prices set the margin. Mistral · AWS · Previously · 29 Oct

• Ordering agents that merchants own, on WhatsApp. The figures from DoorDash and Uber show that outside agents bring almost no orders. The orders that do come through agents arrive on merchants’ own channels: DoorDash’s text-to-order agent and MCP server, Block’s voice ordering for Square sellers, and Pick n Pay’s Penny assistant. In the US these channels are apps, and incumbents build their own. In South Africa ordering already happens on WhatsApp, and small restaurants and retailers cannot build agents themselves. Catch: Meta’s new business terms, due 20 Nov, may change pricing and what bots may do. TechCrunch · PYMNTS on Block · Previously · 29 Oct

Model and API releases

Traction

• FlickReels: AI Story & Drama. This app of AI-made short dramas jumped from 40 to 1 on South Africa’s top-free chart and entered the US list at 40, 29 Oct to 5 Nov. From the Digest’s App Store snapshots. App Store
• Matric study apps. During final exams, Grade 12 Life Sciences went from 7 to 1 on South Africa’s top-paid chart. Its solo publisher, Tshepo Sadiki, holds four of the top 13 places. Private LLM, an on-device chat app, entered at 16. From the Digest’s App Store snapshots, 29 Oct to 4 Nov.
• Assistants on the South African charts. Claude fell from 11 to 26 and ChatGPT from 3 to 5 on South Africa’s top-free chart. In the US, Grok entered the top 50 at 35 and Muse holds 2. From the Digest’s App Store snapshots, 29 Oct to 5 Nov.
• Agentforce Coworker. Salesforce says more than 185,000 people use it, 21 times as many as three months ago, with Adecco and Fulton Bank among the customers that are expanding. Claims · Salesforce
• Repos. thedotmack/claude-mem gained 9,118 stars to 113,268, the fastest growth at 8.8%. deepseek-ai/deepseek-harness added 7,928 and stablyai/orca 6,474. GitHub star counts, 29 Oct and 5 Nov.

Opinions worth reading

• Ed Zitron, Where’s Your Ed At. He argues the money is running out: OpenAI’s shrinking round shows sovereign funds and venture investors have given what they will. It is the week’s sceptic, and it gathers the funding data in one place. wheresyoured.at
• Matthew Honnibal. He argues the viral claim that GPT-6 Astra can read raw weights and edit them to improve Qwen is overstated. The maths gain came from Python programs, not from reading weights alone. Read it before you cite the result. honnibal.dev
• SemiAnalysis. Capacity for 2027 is largely locked in. 2028 is less secure, and a long Texas permit freeze is the biggest single risk to the US build-out. It is the most specific forecast of what the vote changes. semianalysis.com
• Ben Thompson, Stratechery. He asks what it means for AI if US opposition to building data centres is permanent. stratechery.com (paywall)

Builder notes

• If anyone on your team ran Cursor with the Claude Code extension around 30 Oct, check for version 2.1.313. On affected machines, check authorized_keys and scheduled tasks, rotate the credentials those environments held, and turn off automatic extension updates. Ars Technica
• Ubuntu 22.04, 24.04 and 26.04 LTS had no fixed kernel for CVE-2026-98432 (CVSS 9.8) on 31 Oct. Multi-tenant GPU and agent hosts on stock LTS kernels should move to a patched kernel or isolate the hosts until Ubuntu ships a fix. freenode
• Hard spend limits are now generally available on two platforms. Google Cloud budget spend caps reached GA on 30 Oct. Anthropic’s Python SDK 2.4.0 adds GA methods for spend limits and role-based access control. Set caps per workspace before agent workloads grow. Google Cloud · Release
• For personal Google AI subscriptions, saved Gemini chats are used to train models unless Keep Activity is off. Some Codex users report that the “Improve the model for everyone” setting turned itself back on. Do not let staff put customer data into consumer AI accounts, and check data controls after every update. Google · Previously · 29 Oct
• LangChain will drop Python 3.10 in langchain-core 1.8.0 and langchain 1.5.0. Move services to 3.11 or later before the next minor release. Release

Money, markets & policy

Big tech and capital. Microsoft plans to raise US$25bn of debt and Meta up to US$25bn in bonds. That follows Meta’s equity raise last week. The FT reports that Oracle cancelled a US$7bn AI deal with Tencent under pressure from the White House. The FT also reports that Nvidia told cloud providers most of their clients must be non-Chinese. CNBC on Microsoft and on Meta (feed summaries) · FT on Oracle and on Nvidia (paywall, feed summaries) · Previously · 22 Oct

Rules. On 4 Nov the OCC issued a supervisory framework that its examiners will use for banks’ AI, which it now calls “Super Intelligence”. China plans safety standards that cover AI from development to deployment. On 3 Nov an adviser to the EU’s top court said Gemini’s use of publishers’ articles to answer queries may breach copyright. OpenAI, Google, Microsoft and Anthropic will honour a shared opt-out signal that lets publishers refuse AI training. Finextra (feed summary) · Bloomberg (paywall, feed summary) · Reuters via Daily Maverick · The Verge · Previously · 24 Sep

Africa and South Africa. The redrafted national AI policy will not reach Cabinet before March 2027. The drafting team must submit it to the minister by 30 Nov, and internal consultation runs from December to February. Even after March, Cabinet will only be approving it for public comment. OPay, which operates in Nigeria, halted its US$304m NYSE listing hours after pricing it. Bloomberg reports that a Chinese regulator’s inquiry was the cause. The US will remove South Africa from AGOA next year. iAfrica · Bloomberg (paywall, feed summary) · Semafor · Previously · 22 Oct

Radar

• From about 9 Nov. Anthropic’s roadshow, now with Q3 figures and a new warning that it may soon meet its own Responsible Scaling Policy threshold for “dramatically accelerated progress attributable to AI”. TechCrunch · Previously · 29 Oct
• 10 Nov. SABS AI-standards comments close. 20 Nov. Meta’s new WhatsApp business terms, unless they slip again. 24 Nov. Earliest Opus 4.5 retirement. 30 Nov. Deadline for the AI policy draft to reach the minister. Previously · 29 Oct
• 30 Nov to 4 Dec. AWS re:Invent. AWS added in-country inference in India this week. No provider has yet announced frontier inference hosted in an African region. Previously · 29 Oct
• 2 Dec. EU watermarking grace period ends. March 2027. Redrafted national AI policy to Cabinet, and the first non-bank payment participants.

---