OpenAI and Visa let ChatGPT agents make purchases on your behalf. A German court ruled Google liable for false AI Overview statements — saying nobody needs AI to search. And Google DeepMind released DiffusionGemma: 1,000 tokens per second, 4x faster, runs on a consumer GPU. Three stories. One very busy Wednesday.
AI technology updates this week met commerce, law, and model development simultaneously. On 10 June, OpenAI and Visa announced a partnership allowing AI agents to make purchases online after users give their permission. The same day, a German court issued a preliminary ruling holding Google liable for false statements in AI Overviews — the first known time any court held an AI firm liable for AI-generated speech. Meanwhile, Google DeepMind released DiffusionGemma — an experimental open-weight model that generates text four times faster than traditional autoregressive models. All three stories describe AI entering new operational territory simultaneously.
What’s Happening & Why It Matters
ChatGPT and Visa: AI Agents That Shop
OpenAI’s Visa partnership announced at Visa’s Payments Forum in San Francisco marks the most ambitious attempt yet to turn AI agents into genuine commerce participants. ChatGPT will gain the ability to facilitate purchases through Visa’s payment infrastructure — including tokenised transaction processing, real-time fraud prevention, and user-approved spending controls. A user tells ChatGPT to buy groceries, pay a bill, or order a product. The agent handles the transaction. Visa processes the payment. The merchant receives funds as in any standard Visa transaction. Visa chief product and strategy officer Jack Forestell described the ambition directly. “AI will transform commerce more profoundly than the internet or mobile technology ever did.”
OpenAI quietly discontinued its previous Instant Checkout feature in March 2026 — it had never scaled beyond a small number of merchants due to a 4% fee OpenAI charged per sale. The Visa partnership attempts to fix that failure by using Visa‘s existing merchant infrastructure — removing the need for any separate onboarding. Mastercard is developing a competing AI-enabled payment solution — confirming the payment networks see agentic commerce as a strategic priority.
German Court Rules: Nobody Needs AI to Search
The second story is the one that will generate the longest legal tail. A German court issued a preliminary injunction against Google for false statements in its AI Overviews — ruling in a case where two publishers found that AI Overviews incorrectly linked them to scams. The court’s logic is direct and specific. Google tried to argue that most users understand AI outputs aren’t always accurate and must be verified. The court rejected that argument. The court found that, unlike traditional search engines that present lists of links to third-party statements, AI Overviews make statements as if they are Google’s own — and because Google did not correct false statements, it “must be held accountable.”
The ruling’s most striking phrase will travel. “Nobody needs AI to search the internet” — the court’s own language — crystallises the argument that AI Overviews are a product choice, not a necessity. The outlook supports the UK CMA’s publisher opt-out ruling and may influence EU Digital Markets Act enforcement. As TF covered in its UK CMA Google article, publishers across Europe are fighting exactly the same battle.
DiffusionGemma: 1,000 Tokens Per Second on a Single H100
The third story is the most technically significant for developers. Google DeepMind released DiffusionGemma — an experimental open-weight model built on a fundamentally different architecture to every major language model currently available. Rather than generating text one token at a time — left to right, as GPT, Claude, and Gemini all do — DiffusionGemma starts from a canvas of random tokens and refines the entire block in parallel, similar to how image diffusion models generate images from noise.
DiffusionGemma generates text 4x to 5x faster than equivalent autoregressive models on NVIDIA GPUs — achieving over 1,000 tokens per second on a single H100. The model carries 26 billion total parameters as a mixture-of-experts architecture, but activates only 3.8 billion parameters during inference — fitting within the 24GB VRAM of a consumer NVIDIA RTX 5090 or 4090 when quantised. Furthermore, the model integrates with production inference stacks including vLLM, NVIDIA NeMo, Google Cloud Model Garden, and NVIDIA NIM. Google recommends DiffusionGemma for speed-critical workloads — specifically in-line editing and code infilling — rather than applications requiring maximum output quality. It ships under an Apache 2.0 licence, making it free for commercial use.
TF Summary: What’s Next
The ChatGPT-Visa integration rolls out progressively to users who connect their Visa cards to ChatGPT. No specific launch date has been confirmed. The German court’s preliminary ruling requires Google to issue a temporary injunction against spreading the false AI Overview claims — a full hearing follows. Google is expected to appeal. DiffusionGemma weights are downloadable from Hugging Face immediately under Apache 2.0.
MY FORECAST: All three AI technology updates this week will produce lasting changes to the AI competitive landscape. The ChatGPT-Visa integration will succeed where Instant Checkout failed — because Visa‘s merchant network removes the onboarding barrier that killed the previous attempt. By the end of 2026, agentic purchasing will be a standard ChatGPT feature with meaningful transaction volume. The German court ruling will produce copycat cases across Europe within months — publishers in France, Spain, and the Netherlands will cite it as precedent. Google faces a wave of AI Overview liability claims that the UK CMA ruling, the German injunction, and the EU AI Act’s accuracy requirements collectively produce. And DiffusionGemma’s 4x speed advantage will make it the default choice for real-time code completion products within 12 months — not because it is the smartest model, but because at 1,000 tokens per second, developers can stop waiting.