OpenAI puts ads in your AI images and DeepSeek closes the gap to 3%

Monday's news all points the same way: the AI boom is being asked to pay for itself. OpenAI is selling your attention, DeepSeek is selling the same capability for less, and the people who built the chips are suing over who got paid.

OpenAI will put visual ads next to the images you generate

Later in October OpenAI starts testing image ads during ChatGPT's image generation, in the United States only and with a selected group of advertisers. The ads will be "clearly labeled and kept separate from the images ChatGPT creates," and reach a product with 1.2 billion weekly users.
The same update expands measurement: integrations with Kochava and ten other partners, attribution beyond last click, geo-incrementality tests with Haus, Measured and WorkMagic, and brand-safety pilots with DoubleVerify and Integral Ad Science that never see the actual conversations. The constraint is inventory — roughly 2.5 billion prompts a day against Google's 15 billion searches, with a single ad slot per answer.

Why it mattersOpenAI launched ad sales in February and advertisers complained they couldn't measure anything. This is the fix, and it quietly settles a bigger question: the free tier of the most used AI product in the world is now an ad business, not a loss leader. For anyone selling in Latin America, note the order of events — the format arrives first in the US, the measurement tooling is global, and the day the inventory opens up the cheapest attention in Spanish will be gone fast. If you sell online, start writing down which prompts bring you customers now, before the auction decides the price.

DeepSeek V4.1 Flash cuts the US lead to under 3%

In LiveBench's October 4 snapshot, DeepSeek V4.1 Flash at max effort scored 81.1 against 83.4 for Anthropic's Claude Fable 5.1 — a gap of about 2.8%, down from roughly 9% in May. On agentic coding the Chinese model came out ahead: 77.3 versus 66.1.
The gain comes from architecture, not more compute. DeepSeek's September 10 update activates only part of the model's parameters per answer and shrinks the conversation cache, which the company says cuts the computing and memory needed to serve a response.

Why it mattersA 3% benchmark gap is not the same as 3% of the market — US models still dominate enterprise workloads, and Bloomberg Intelligence expects China's AI industry to stay unprofitable until 2030. But the price curve matters more than the leaderboard. For a startup in Caracas, Bogotá or Mexico City paying per token, a model that is 97% as good and a fraction of the cost is the one that gets built on. The practical move is to stop assuming the frontier model is the right default and benchmark your own task against a cheap one — the gap you care about is the gap on your use case, not on LiveBench.

Ex-Groq engineers sue the board over Nvidia's $20 billion deal

Joshua Rubin and Benjamin Serebrin filed suit in the Delaware Court of Chancery on October 2 over the structure of Nvidia's deal: $17 billion for a non-exclusive license to Groq's inference technology and $3 billion in restricted stock for the engineers who left for Nvidia. They allege the board skipped a required stockholder vote, failed to maximize the value of the assets, and had conflicts of interest, with investment funds on the board positioned to profit.
The deal was announced in December; founder Jonathan Ross and president Sunny Madra went to Nvidia along with 150 to 200 engineers, and the resulting Groq 3 LPX inference chip entered full production in 2026. Groq has since raised about $1 billion and calls the suit "meritless," saying the license "delivered exceptional value for Groq, our investors, and our employees." The DOJ opened its own probe in September.

Why it mattersThis is the acqui-license taking its first real legal hit. Buying the technology and the team without buying the company lets a giant skip antitrust review and skip the shareholder vote — and leaves the shareholders who funded the research holding a company whose crown jewels are now licensed to its biggest competitor. If Delaware says that structure needs a vote, every deal built this way in the last two years gets re-read. Founders anywhere should read the lesson narrowly and usefully: the clause that decides who gets paid is the one defining what counts as a sale of substantially all assets.

California will fine robotaxis $10,000 for blocking emergency crews

Governor Gavin Newsom signed SB 1246, which fines operators up to $10,000 per driverless vehicle that blocks emergency responders for more than 30 minutes. It also requires local incident technicians on site, US-based remote drivers holding valid licenses, notification of authorities during system failures, and reporting of response times.
The law takes effect July 1, 2028. The record behind it: Waymo vehicles stranded after San Francisco's December 2025 power outage, a Waymo car blocking an ambulance responding to a shooting in Austin, and NHTSA describing "a clear pattern of driverless AVs interfering with law enforcement." Waymo has acknowledged that the humans who unstick its cars are sometimes located overseas.

Why it mattersThe fine is small for a company like Waymo; the staffing requirements are not. Demanding a licensed remote operator physically inside the United States plus a technician who can reach the vehicle turns autonomy into a payroll problem, and payroll is what autonomy was supposed to remove. Watch the offshore-operator clause — it is the first law to say explicitly that the human behind the robot cannot be in another country. Any Latin American city negotiating with an AV company should ask for the same two things California just made mandatory: who shows up, and in how long.

Jeeves raises $110M to run corporate money on stablecoins

The Mexico-founded fintech announced a $110 million equity round on October 1, led by CoinFund with AllianceBernstein, Andreessen Horowitz, Coinbase Ventures, CRV, GIC, ParaFi, Vista, Wintermute and Y Combinator participating. CEO Dileep Thazhmon's platform bundles corporate cards, cross-border payments, accounts payable and treasury on stablecoin settlement.
Annualized platform volume is above $5 billion, tripled year over year, with roughly $1.5 billion a year settling on-chain. More than 80% of customers use more than one Jeeves product, and the company is pushing toward 35 countries — recent additions include Argentina, Costa Rica, the Dominican Republic, Guatemala, Panama, Peru, Paraguay and Uruguay, plus a new Madrid office.

Why it mattersThis was the biggest Latin American round of the week by a wide margin, and the thesis is regional, not crypto-speculative: the stablecoin is plumbing for companies that need to pay a supplier in another country on a Friday afternoon. For Venezuelan and LatAm businesses the useful signal is that this plumbing is becoming boring, institutional and funded by a16z and GIC — which is exactly when it starts being usable by a company that is not a startup. If your business bills in dollars from the region, the number worth tracking is settlement time, and it is the one this round is paying to shrink.

The big picture

Four of these five stories are about the bill arriving. OpenAI needs the free tier to pay rent, DeepSeek is undercutting the price of the same capability, Groq's shareholders want their share of a $20 billion transfer, and California is putting a dollar figure on an autonomous car that stops moving. The era where AI progress was measured only in capability is ending; from here it gets measured in who pays, who collects, and who is liable when it fails.

Go deeper

Sources & further reading

Topics:AI Chips Regulation Startups