Seven days, two concrete shifts. On one side, the AI agent stops advising and starts acting: Robinhood opens share trading and card payments to it, Google entrusts its AI with the full running of display advertising, Meta puts its tools behind a subscription. On the other, the narrative is being renegotiated at the top: Sam Altman and Dario Amodei walk back their prophecies of a jobs collapse a few months before IPOs valued at around a thousand billion dollars, while an Anthropic co-founder warns the Vatican about the labs' conflicts of interest, and a super PAC spends millions to bring down a pro-regulation elected official — funded by part of the industry, fought by the other. In the background, DeepSeek makes a 75% price cut permanent, Huawei announces a route around Western etching machines, and Microsoft returns to the image race. The week agentic AI touched real money while its leaders adjusted the story.
In this issue
- Robinhood opens trading and payments to AI agents — FINRA sounds the alarm
- Altman and Amodei retract their jobs predictions, ahead of their IPOs
- Chris Olah at the Vatican: an Anthropic co-founder names the conflicts of interest
- "Leading the Future" against Alex Bores — and Anthropic funds the other side
- SpaceX heads for the Nasdaq on 12 June — the Tesla merger returns, after the one with xAI
- DeepSeek makes the 75% cut on V4-Pro permanent
- Huawei unveils "LogicFolding" to get around EUV
- Microsoft MAI-Image-2.5 enters the top 3 of image generation
- Meta launches "Meta One": AI goes behind a toll at $7.99
- Google folds the Display Network into Demand Gen
- What to take away
- FAQ
1. Robinhood opens trading and payments to AI agents — FINRA sounds the alarm
On 27 May 2026, Robinhood announced that it would let its customers connect an AI agent to a separate investment account via the MCP protocol (Model Context Protocol). The agent — Claude, ChatGPT, Codex or Cursor, as long as it can speak MCP — can read the account value, buying power, positions and order history, then buy and sell shares autonomously, and even make credit card purchases (TechCrunch, 27 May 2026, The Decoder).
FINRA, the US brokerage self-regulatory authority, has classed AI agents as a new risk area in its 2026 oversight report: an agent can act without human validation, exceed the user's intent, make decisions that are hard to reconstruct, or accidentally expose sensitive data. It recommends that brokers trace agents' actions and identify the points where human oversight remains necessary. American Banker called the launch a "wake-up call" for banks.
For operators. The hard part is not technical, it is legal. Neither the SEC nor FINRA has a written rule for the case of an order placed by an agent — erroneous, unsolicited, or simply misunderstood. The question of liability (the user, the broker, or the model's publisher?) remains open, and it is precisely this void that Robinhood is testing at full scale. Any company considering delegating actions with financial consequences to an agent should first write down in black and white who answers for an error, before plugging in MCP.
2. Altman and Amodei retract their jobs predictions, ahead of their IPOs
Two reversals in the same week. Sam Altman, who warned in June 2025 about the threat to junior positions, said in an interview with the chief executive of the Commonwealth Bank of Australia that he had been "pretty wrong" about the economic impact of AI: "I'm delighted to have been wrong about that". Dario Amodei, who argued that AI could eliminate up to 50% of office jobs, now presents automation as a productivity multiplier: automating 90% of a job would, in his view, make the remaining 10% grow until productivity is multiplied tenfold (Fortune, 26 May 2026, Time).
The timing feeds the doubt. Both companies are preparing an initial public offering this year, each valued at around $1,000 billion — an order of magnitude unrelated to the "$380 billion" figure circulating in some summaries, which corresponds to a private mark already surpassed. A narrative of job-market collapse does not help to woo cautious institutional investors.
3. Chris Olah at the Vatican: an Anthropic co-founder names the conflicts of interest
On 25 May 2026, Pope Leo XIV published an encyclical on AI, "Magnifica humanitas", devoted to safeguarding the human person in the face of artificial intelligence. At its presentation, Anthropic co-founder Chris Olah made remarks that, by his own admission, "may seem strange coming from the co-founder of an AI company": every frontier lab, Anthropic included, "operates within a set of incentives and constraints that can sometimes conflict with doing what is right" (Anthropic, Futurism, May 2026).
Olah made two observations. First, work: a real possibility of labour displacement on a very large scale, the management of which would be "a moral imperative of historic proportions" — the very prediction that Altman and Amodei have just softened (see section 2). Second, concentration: the development of AI is spread among a handful of wealthy nations, and "we have no mechanism" to share its gains globally, an unresolved problem. Finally he mentioned his team's work on the inside of the models, which "keeps finding mysterious, even troubling things": structures that echo human neuroscience, signs of introspection, internal states that functionally mimic joy, satisfaction, fear, grief and unease.
Setting this beside section 2 speaks for itself. In the same month, two lab leaders reassure the market about jobs; the co-founder of a third, in front of the Vatican, calls that displacement a moral imperative and explicitly names the conflicts of interest in his own sector. This is not a contradiction of the ecosystem — it is, more modestly, two registers of speech addressed to two different audiences.
4. "Leading the Future" against Alex Bores — and Anthropic funds the other side
A pro-AI super PAC called "Leading the Future" is spending millions to defeat the congressional candidacy of Alex Bores, a New York Assembly member. His offence in the eyes of part of Silicon Valley: having carried the RAISE Act, a transparency law signed in December, which requires labs with more than $500 million in revenue to publish a safety plan, stick to it and report any serious incident (TechCrunch, 3 March 2026, TechCrunch, 17 Nov. 2025).
The PAC, formed in August 2025 with a commitment of more than $100 million, raised $125 million in 2025 and carried $70 million over into 2026. Its backers include Andreessen Horowitz (a16z), Greg Brockman (OpenAI's president, in a personal capacity), Joe Lonsdale (Palantir co-founder, via 8VC) and Perplexity. This is where the quick summary misleads: it is not a case of "OpenAI, Palantir and a16z" as companies forming a bloc, but of a coalition of funds and individual executives.
The decisive nuance lies in the other camp. Anthropic gave $20 million to Public First Action in February 2026, on the pro-regulation side. The AI industry is therefore not united against oversight: it is divided, and funds both sides. As for the operation itself, it produced a textbook Streisand effect — the campaign made Bores the national face of AI regulation, and he replied "bring it on".
5. SpaceX heads for the Nasdaq on 12 June — the Tesla merger returns, after the one with xAI
SpaceX, whose S-1 was filed on 20 May, is expected to start trading on the Nasdaq around 12 June under the ticker SPCX, with a targeted valuation of up to $1,750 billion and a raise of around $75 billion — the largest IPO in history. A detail that many summaries omit: this valuation rests on a base of $1,250 billion obtained after the recent merger of SpaceX with xAI, Elon Musk's AI company. AI is not adjacent to the deal, it is already within its scope (CNBC, 26 May 2026, Semafor).
Against this backdrop, speculation about a merger with Tesla has resumed. Elon Musk has floated the idea to those close to him; Dan Ives (Wedbush) now puts its probability at 80-90%. Lawyers believe such a transaction would raise mostly shareholder questions, few antitrust ones — and Musk controls 85% of the voting rights at SpaceX. Electrek sees in it the billionaire's "fourth self-deal" (Electrek, 27 May 2026).
For operators. The AI interest of such a rapprochement is not the stated megalomania, but concentration: Tesla is pushing its Dojo supercomputer and its inference chips for driving, SpaceX already integrates xAI. A Tesla-SpaceX-xAI group would bring together compute, driving data and space data under a single governance. It is a signal to follow for anyone who depends, directly or not, on this ecosystem's compute capabilities.
6. DeepSeek makes the 75% cut on V4-Pro permanent
DeepSeek has confirmed that the 75% price cut on its V4-Pro model is becoming permanent, not promotional (The Register, ExplainX). The cut rests on a Hybrid Attention architecture which, compared with V3.2, reduces the KV cache memory footprint by 90% and per-token inference compute by 73% — a combination of Compressed Sparse Attention and a "Lightning Indexer" selector in FP4.
Precision matters: the -73% concerns compute, the -90% memory, the -75% the price charged. These are three distinct figures that a hasty reading merges into one. V4-Pro is now advertised at an input cost of the order of a few thousandths of a dollar per million cached tokens, or, according to DeepSeek, up to twenty times cheaper than a comparable proprietary model.
For operators. This cut confirms a trend already measured on the developer side: on long-context workloads, the price gap between Chinese open models and proprietary models is widening faster than it is narrowing. For a high-volume product, routing part of the non-sensitive traffic to this type of model is no longer a gamble, it is a margin line — provided you have settled, upstream, the question of the confidentiality of the data sent.
7. Huawei unveils "LogicFolding" to get around EUV
At the ISCAS 2026 symposium in Shanghai (25-26 May), Huawei's head of semiconductors, He Tingbo, presented an architecture called "LogicFolding". Rather than shrinking transistors — which requires EUV etching machines that China cannot access — the approach consists of folding 2D circuits into 3D vertical structures, by stacking layers (Fortune, 25 May 2026, Tom's Hardware).
Huawei claims a 55% increase in transistor density and 41% energy efficiency, and proposes a "Tau law" meant to take over from Moore's law by optimising signal travel time rather than etching fineness. The stated goal: to produce 1.4 nm-class chips by 2031. To keep in mind as a counterweight to the announcement effect: TSMC is aiming for the same class as early as 2028. The gap is still measured in years, not months — but it now has a name, where it was presented as insurmountable.
8. Microsoft MAI-Image-2.5 enters the top 3 of image generation
In late May, Microsoft's MAI team released MAI-Image-2.5, its new text-to-image model, which ranks third on the Arena leaderboard with an overall score of 1254. The podium: OpenAI's gpt-image-2 in first place (≈1388), Google's Gemini 3.1 Flash Image second, MAI-Image-2.5 third — the "OpenAI's Image-2" mentioned in summaries in fact refers to gpt-image-2 (Microsoft AI, JustGeek, Neowin).
The detail of the sub-scores sheds light on its positioning: MAI-Image-2.5 is particularly strong on text rendering (1278), stylised illustration — cartoon, anime, fantasy (1268) — and commercial visuals, packaging and branding (1263), three historic weaknesses of image generation; it remains further behind on pure photorealism (1247). In practical terms, its playground is usable brand visuals (legible typography, product mock-ups) rather than photography. Microsoft announces an improvement across all the categories assessed compared with MAI-Image-2. The model is already available on Arena, and will arrive on the MAI Playground and Azure Foundry within two weeks.
9. Meta launches "Meta One": AI goes behind a toll at $7.99
Meta is testing two subscriptions to its AI services under the "Meta One" brand: a Plus plan at $7.99 a month for intensive use of image generation, video and extended reasoning, and a Premium plan at $19.99 opening up more compute capacity (CNBC, 27 May 2026, TechCrunch). The test starts next month in Singapore, Guatemala and Bolivia; Meta AI remains free for occasional use. It is a turning point: tools that have so far been free are entering a logic of direct monetisation, as the cost of the compute behind Meta's "superintelligence" begins to bite.
10. Google folds the Display Network into Demand Gen
Google is retiring standalone Display campaigns in favour of its AI-driven Demand Gen platform. From June 2026, a migration tool will appear in Google Ads; after that, it will no longer be possible to create new independent Display campaigns, with the process stretching into 2027 (Search Engine Journal, blog.google). Display inventory does not disappear: it joins Demand Gen alongside YouTube, Discover, Gmail and Maps.
The change of method runs deeper than the change of brand. The advertiser no longer chooses sites or adjusts audience segments: they supply a business objective and raw creative assets — images, clips, hooks — and the AI decides on the combinations, formats and placements. Manual control hands over to predictive automation on one of the world's largest advertising networks.
11. What to take away
Five points from the week, phrased for action rather than for conclusion.
| Theme | Move | Immediate consequence |
|---|---|---|
| Agentic finance | Robinhood opens trading and payments to agents via MCP; FINRA classes the risk | Settle liability (who answers for an erroneous order?) before delegating any financial action to an agent. |
| Jobs narrative | Altman and Amodei soften their predictions ahead of IPOs at ~$1,000bn; Olah contradicts them at the Vatican | Do not mistake a leader's statement for an independent assessment of the risk. |
| Regulation | "Leading the Future" targets Bores (RAISE Act); Anthropic funds the opposing camp ($20M) | The industry is divided, not united. Read each position in light of who funds it. |
| Costs | DeepSeek makes -75% price permanent on V4-Pro (-90% KV cache, -73% FLOPs) | Routing non-sensitive traffic to open models becomes a margin line, subject to confidentiality. |
| Advertising | Google folds the Display Network into Demand Gen, fully automated | Take back control of creative assets and measurement, the only levers left when targeting becomes a black box. |
First, the AI agent crossed the frontier of real money this week — at Robinhood on the markets, at Google on advertising budgets. The practical question is no longer "what can the agent do?" but "who answers when it gets it wrong?", and nobody has yet answered it in writing.
Second, the gap between the labs' registers of speech deserves to be held for what it is: a fact observed over a few days, not a theory of the ecosystem. Altman and Amodei reassure the market, Olah alerts the Vatican. Rather than a conclusion, a discipline of reading: relate each statement to its financial context and its audience.
Third, the economics of inference continues to tilt towards open models, and compute sovereignty is becoming a concrete industrial subject again — DeepSeek on price, Huawei on manufacturing. Neither of these movements concludes within a week; both will be verified over the trajectory of the coming quarters.
12. FAQ
What exactly does Robinhood allow with AI agents?
Robinhood lets you connect an AI agent (Claude, ChatGPT, Codex, Cursor) to a separate investment account via the MCP protocol. The agent can read the account value, buying power, positions and order history, then buy and sell shares autonomously, as well as make card payments. FINRA has classed AI agents as a new risk area in its 2026 oversight report, in the absence of dedicated SEC or FINRA rules for this type of automation.
Why are Sam Altman and Dario Amodei walking back their predictions on jobs?
Both leaders had predicted massive job losses in 2025 — junior positions under threat for Altman, up to 50% of office jobs for Amodei. In May 2026, Altman says he was "pretty wrong" and Amodei presents automation as a productivity multiplier. The reversal comes as both companies prepare initial public offerings valued at around $1,000 billion each. One possible reading: a narrative of job-market collapse is not a selling point in front of cautious institutional investors.
What is the "Leading the Future" super PAC and who funds it?
Leading the Future is a pro-AI super PAC formed in August 2025 with a commitment of more than $100 million. It raised $125 million in 2025 and carried $70 million over into 2026. Its backers include Andreessen Horowitz (a16z), Greg Brockman (OpenAI's president, in a personal capacity), Joe Lonsdale (Palantir co-founder, via 8VC) and Perplexity. Its first target is Alex Bores, author of the RAISE Act, a transparency law signed in December. By contrast, Anthropic gave $20 million to Public First Action in February 2026: the industry is divided on regulation, not united.
Where does SpaceX's initial public offering stand?
SpaceX filed its S-1 on 20 May 2026 and is expected to start trading on the Nasdaq around 12 June under the ticker SPCX, with a targeted valuation of up to $1,750 billion and a raise of around $75 billion — the largest IPO in history. This valuation rests on a base of $1,250 billion obtained after the recent merger of SpaceX with xAI. Speculation about a later merger with Tesla has resumed: Dan Ives (Wedbush) puts its probability at 80-90%.
What weak signals should we watch next week?
Several deadlines are close: SpaceX's roadshow and trading debut around 12 June on the Nasdaq; the migration of Google Display campaigns to Demand Gen starting in June; the trial of Meta One subscriptions starting in Singapore, Guatemala and Bolivia; and the wider availability of Microsoft's MAI-Image-2.5 on Foundry within two weeks.
This note is the third recap in a weekly series published every Friday. The next issue is out on 5 June 2026.