
OpenAI supercharges ChatGPT and launches new agentic reasoning models o3 and o4-mini.
Google DeepMind releases Gemini oracles Virtual Open House: Gemini 2.5 Flash, a “hybrid reasoning” LLM that gives devs the ability to “dial in” thinking on or off for cost-latency tradeoffs.
NVIDIA begins U.S. manufacturing of Blackwell-based AI super-computers as it lobbies Washington to relax export restrictions.
EU bets €200 billion on becoming the “AI Continent” and begins consultations on its Cloud & AI Development Act.
Visa bets on agentic checkout
Visa’s new Intelligent Commerce platform will allow consumers to hand off the entire shopping flow — search, compare, book and pay — to AI agents created with Microsoft, OpenAI, Anthropic, IBM, Mistral AI, Perplexity, Samsung and Stripe. Shoppers maintain their hard spending cap; the agent does the legwork and provides one single “approve or reject” screen. Visa’s pitch: reduce abandonment and catch impulse buys on the spot in sub-30-second sessions — the break-even point where 1 in 3 currently bails.
Why it matters:
Agentic commerce pushes the industry past chatbots into full-on automated consumer AI. Payments rails that solve the “last click” problem will be sticky integration points for retailers, BNPL lenders, and ad networks. Look forward (or not) to Shopify, PayPal, and Amazon to respond—or get completely disintermediated.
OpenAI further prepares for retail & reasonotent.
Two distinct OpenAI drops this week:
Shopping search: ChatGPT now offers you product cards — images, prices, reviews, buy-links — with zero advertisements, in the fashion, beauty and electronics domain. Available to free users, and even anonymous sessions. The easy target: Google’s search ads powered by PLAs.
o-series refresh: o3 and smaller o4-mini o3 and the lighter-size model o4-mini continue “let the model think” philosophy with tooling (web, Python, vision, image-gen) on board. Initial benchmarks are showing ~20% fewer major errors than o1 and the state-of-the-art performance on SWE-bench and Codeforces.
Strategic take: OpenAI is collapsing search, commerce, and agentic tool-use into one surface. If the product cards are conversions without the ads, the revenue model moves to premium tiers and enterprise APIs — endangering classic CPC search income for Google.
DeepMind’s Gemini 2.5 Flash – Throwing money at the problem or thinking out of the box?
Google’s is Gemini 2.5 Flash, a “fully hybrid reasoning” model in which developers can establish a “thinking budget” – essentially metering the depth of the chain of thought in order to tune cost and latency. And this capable behemoth easily outpaces 2.0 Flash even with thinking disabled; competing even with larger models at a fraction of the compute at modest budget.
Futurwise: LLM APIs will offer configurable cognition tiers (“day-trader mode”, “deep analyst mode”) as cloud instances advertise CPU vs. GPU hours today.
NVIDIA remakes the AI supply chain
Made-in-USA silicon: Blackwell GPU packaging begins in Arizona; complete AI-supercomputer assembly in Texas with Foxconn & Wistron. Goal: U.S. half-trillion-dollar AI infra over four years.
Policy pressure: CEO Jensen Huang went public calling on the Trump administration to dial back strict export bans, saying gating China too hard only stimulates local competition and crimps U.S. revenue.
Impact: Domestic fabs create geopolitical choke point hedge and get CHIPS Act incentives, but export-license friction could dampen near-term sales. Keep an eye on semiconductor ETFs for the potential for volatility as policy signals emerge.
The European Commission announced a €200 billion investment blueprint, covering compute infrastructure, data access, and sector-specific adoption incentives — all underpinned by compliance with the final EU AI Act. Public consultations on the Cloud & AI Development Act continue until 4 June.
Why it matters: Brussels is positioning to be a builder, not just a regulator. Anticipate funding pipelines for sovereign GPU clusters (Gaia-X 2.0), open-source model grants and heavy demand-pull health-tech & manufacturing. Vendors based in the United States and Asia will vie for E.U. projects but have to navigate strict transparency and risk-classification regulations.
Quick hits & signals to watch
In October, Anthropic released Claude for Education and new Workspace-integrated research mode—announcing a land grab in academia with the neophile vertical strip.
Checkpoint drought open source: No major news this month; OSS communities are eagerly waiting to see how a policy changes (in particular the European Union) will disrupt model licensing.
Coming up next: Google I/O (May 14) – when we’ll likely see Gemini’s recently introduced video funnel generator Veo 2 and a Gemini 3 preview.
The Bottom Line
Agentic workflows went from demo to deployment this week — on payments, on search, and with enterprise tooling — as infrastructure and policy jockeyed to keep pace. For companies, the short-term play is integration: baking these agentic APIs in while the user experience is still gelatinous enough to prevent consolidation around a few dominant platforms. For investors, monitor the localization of GPU supply chains, policy-risk hedges, and the chip revenue race between ad-funded search and subscription-based AI assistants.
Stay tuned — next week’s Pulse will dissect I/O announcements and any surprise filings in the AI Regulation pipeline.
Leave a comment