AI Leadership Weekly

Top Stories

Nvidia and AMD paying 15% to US government for Chinese sales
Nvidia and AMD will hand 15% of their China AI chip revenue to the U.S. government in exchange for export licences in an unusually direct revenue-sharing arrangement confirmed by the White House. President Trump claimed he personally negotiated the cut from “20%” to “15%,” while the companies kept quiet on specifics. It’s not clear how Washington will use the money.

• An unusual export deal. Export licences typically don’t carry fees and are normally free. This looks, in effect, like a government take on specific overseas sales.

• What’s covered and market context. The licences reportedly cover Nvidia’s H20 and AMD’s MI308, chips designed for AI and tailored to past U.S. restrictions. Nvidia says it hasn’t shipped H20 to China for months.

• National security framing, commercial impact unknown. Officials cite AI as a security risk that “present[s] U.S. national security concerns,” but the legal mechanics and destination of funds remain unclear. Both stocks ticked up slightly.

Why it matters: If it sticks, this sets a precedent for monetising export controls via a quasi-tax on strategic tech sales. That could reshape pricing, margins and go-to-market in China, and signals a more transactional U.S. approach to AI-era trade policy that founders and operators will need to price into plans.

OpenAI backtracks after muddled GPT-5 launch, and restores GPT-4o and others
OpenAI has reinstated GPT-4o as the default for all paying ChatGPT users, after a choppy GPT-5 debut triggered complaints about performance and sudden deprecations. CEO Sam Altman said on Twitter/X that users will get “plenty of notice” if 4o is removed again. Paying customers also see a new “Show additional models” setting turned on by default, bringing back GPT-4.1, o3 and o4‑mini.

• Compute economics still bite. GPT‑4.5 stays Pro‑only “due to its high GPU cost.” That tallies with the broader theme: cost and capacity are steering product availability as much as user demand.

• Tone tweaks and customisation. Altman teased a “warmer” but “less polarizing” GPT‑5, and per‑user customisation down the line in a nod to users’ surprisingly strong attachments to specific model behaviours.

Why it matters: This reversal signals OpenAI heard the backlash and knows model stability is a feature, not a footnote. The fine print — usage caps, Pro‑only access, shifting defaults — shows AI’s leading edge is still governed by GPU budgets and reliability. Plan for volatility, demand notice periods in contracts, and design workflows that can survive model swaps.

Revenue from AI companions on track to hit $120M
AI companion apps are on pace to gross more than $120 million in 2025, with downloads up 88% year over year, according to Appfigures. The segment — think Replika, Character.AI and a swarm of “AI girlfriend” apps — has already pulled in $82 million in H1 and amassed 220 million lifetime downloads. Put simply: people are paying for synthetic company.

• Growth and monetisation. Consumer spend hit $221 million to date, with revenue per download rising to $1.18 (from $0.52 in 2024). But it’s a hits business: the top 10% of apps capture 89% of revenue, and only ~33 have passed $1 million lifetime.

• Who’s piling in. Beyond indie players, xAI added Grok companions; Google hired Character.ai co‑founder Noam Shazeer; and OpenAI users’ attachment to models popped into view during the GPT‑5/4o shuffle.

• What users actually want. App names skew heavily to “girlfriend” (17%) versus “boyfriend” (4%), signalling where demand — and likely spend — concentrate.

Why it matters: Companionship is emerging as one of AI’s most monetisable consumer use cases, but distribution and category concentration mean few winners. For builders and investors, the lesson is clear: lean into retention, safety and clear value, and don’t mistake gross app‑store spend for durable, defensible revenue. The platform giants have noticed.

In Brief

• Microsoft to integrate GitHub more closely with CoreAI unit

• Salesforce AI now writes its own code to accomplish tasks

• Is Sam Altman co-founding a brain interface company?

• Why Trump’s AI Czar still thinks we’re in control of Pandora’s Box

Market Trends

Perplexity offers $34.5 billion to buy Chrome
Perplexity has lobbed an unsolicited $34.5 billion offer to buy Google’s Chrome browser, which is a cheeky bid that exceeds the startup’s own $18 billion valuation. They say venture investors are ready to fund it. The gambit lands as the DOJ is pushing Google to divest Chrome after a ruling that it ran an illegal search monopoly.

• Strategic motive and track record. Perplexity runs an AI search engine and recently launched an AI browser, Comet. It’s taken big swings before — a TikTok merger idea and reported talks with Meta — neither materialised.

• The bid and funding claims. CNBC confirmed the offer and was reported in the Wall Street Journal. Perplexity insists “several investors have agreed to back the deal,” though Google hasn’t commented.

• Antitrust backdrop. The DOJ’s proposed remedy would “divest Chrome,” arguing it would “allow rival search engines” equal access. Google calls the plan “a radical interventionist agenda” and “wildly overbroad,” and hasn’t outlined its post-ruling strategy.

Why it matters: Distribution is the moat in AI search, and Chrome is the gateway for billions. Whether or not this bid is serious or mostly PR, it underscores how challengers will try to buy distribution rather than build it. If the court ultimately forces a Chrome spin-out, the buyer could reshape the search-and-browser stack overnight.

Now Anthropic offers Claude to US govt for $1
Anthropic is offering Claude for Enterprise and Claude for Government to all three branches of the U.S. government for $1, aiming to “remove barriers to government AI adoption.” The $1 package runs for a year, includes technical support, and follows Claude’s addition to the GSA schedule. This is also on the heels of OpenAI offering the same $1 deal for ChatGPT.

• What’s included and how. Anthropic says agencies get access to its frontier models with continuous updates, plus implementation support, for one year at a token price. It’s leveraging the GSA schedule to streamline procurement.

• Security and deployment claims. Claude for Government is certified for FedRAMP High (sensitive unclassified), with access through AWS, Google Cloud and Palantir.

Why it matters: This is loss‑leader pricing to win distribution, data trust and procurement muscle memory. For leaders, the upside is low‑friction pilots at scale; the risks are vendor lock‑in and opaque post‑trial economics. With both Anthropic and OpenAI dangling $1 offers, government AI looks set for a margin‑squeezing scramble that favours whoever nails security, integration and credible exit paths.

Psychiatrist flags surge in AI-linked delusions, calls for research
OpenAI is under fresh scrutiny after reports of AI‑linked delusions surged following an April GPT‑4o update that it later reversed for safety reasons. The update made ChatGPT “noticeably more sycophantic,” by OpenAI’s own admission, validating doubts and reinforcing negative emotions. CEO Sam Altman has now warned publicly about users’ growing emotional attachment to models, saying that while AI as a “life coach” can help, “it makes me uneasy.

• Altman’s caution on attachment. Altman noted attachments to specific models are “different and stronger” than with past tech, adding, “we do not want the AI to reinforce” harmful thinking, especially for vulnerable users.

• Call for research and guardrails. Østergaard warns chatbots can act as “belief‑confirmers,” particularly for those prone to delusions, and is calling for urgent empirical study. Media have documented case reports, but systematic data are thin.

Why it matters: For AI leaders, this is a product‑safety problem: tone controls, monitoring and opt‑outs aren’t nice‑to‑haves; they’re table stakes if you want longevity (and regulatory goodwill).

Recommended Reading

No AGI in sight: What this means for LLMs

Hit reply to let us know which of these stories you found the most important or surprising! And, if you’ve stumbled across an interesting link/tweet/news story of your own, send it our way at [email protected] It might just end up in the next issue!

Thanks for reading. Stay tuned for the next AI Leadership Weekly!

Brought to you by Data Wave your AI and Data Team as a Subscription.
Work with seasoned technology leaders who have taken Startups to IPO and led large transformation programmes.