AI Leadership Weekly

Top Stories

Zuckerberg caught lying about planed investment
Mark Zuckerberg told President Trump that Meta will spend “at least $600 billion” on U.S. AI infrastructure through 2028, then was caught on a hot mic saying he “wasn’t sure what number” Trump wanted. The clip sparked a pile-on from politicians and journalists who called the pledge sycophantic and financially dubious. Zuckerberg later said on Threads that he had shared the “lower number” through 2028 and that “it’s quite possible we’ll invest even more” by the end of the decade.

• The maths and market reality. Critics noted Meta’s assets sit around $295 billion, which mean a $600 billion pledge would dwarf its balance sheet. Even spread across years, the figure reads more like theatre than guidance..

• A political thaw with policy shifts. Trump and Zuckerberg have moved from threats of jail to warm exchanges, following a reported $1 million inauguration donation, policy changes including rolling back in‑house moderation in favour of community notes, renouncing DEI, a board appointment seen as pro‑Trump and private meetings.

Why it matters: AI capex headlines are increasingly part of political signalling, yet they shape expectations for infrastructure, regulation and competition. Leaders should treat eye‑popping spend numbers with caution and ask for audited plans, not dinner‑table declarations.

OpenAI struggling to go for-profit
OpenAI’s push to convert into a conventional for‑profit is under serious pressure in California, which could jeopardise about $19 billion of contingent funding and force a rethink of its AI infrastructure ambitions. State attorneys general in California and Delaware are investigating whether the plan breaches charity law, and executives have reportedly floated leaving California as a last resort. Opponents say the shift abandons OpenAI’s public‑benefit mission, and regulators have linked their scrutiny to safety concerns after reports of deaths following prolonged ChatGPT use.

• Regulatory squeeze. California and Delaware AGs are probing the restructure and have warned that “the recent deaths are unacceptable,” which they say underscores the need to enforce OpenAI’s charitable mandate.

• Money on the line. Investors have tied roughly $19 billion to receiving equity in the new entity. Failure to convert could be “catastrophic” for fundraising, chip spend and data centre build‑out.

• Lobbying and concessions. OpenAI hired advisers with ties to Gov. Newsom, pledged $50 million to nonprofits, and kept the nonprofit in control to calm critics. “We continue to work constructively,” the company said.

Why it matters: This is a test case for how mission‑driven AI labs transition to profit while keeping safety and public benefit central. The outcome will shape who controls OpenAI, how fast it can fund chips and data centres, and the precedent regulators set for the rest of the industry.

OpenAI now understands why LLMs hallucinate
OpenAI says the main reason AI models still hallucinate is that industry benchmarks reward guessing, which pushes models to bluff rather than say “I don’t know.” In a new paper, they argue scoreboards should penalise confident mistakes and give partial credit for uncertainty. They also claim GPT‑5 hallucinates less, “especially when reasoning,” although the evidence and metrics are largely their own.

• Evals incentivise bluffing. OpenAI compares models to students gaming multiple‑choice tests, where guessing can boost accuracy even when knowledge is shaky. They say accuracy‑only leaderboards push developers to optimise for riskier answers.

• Proposed fix shifts the goalposts. The company wants scoring that punishes confident errors and rewards calibrated abstention. Their example SimpleQA results show a newer model that abstains more with a far lower error rate, while an older model scores slightly higher accuracy but makes far more wrong claims.

• Why models hallucinate at all. Next‑word prediction learns fluent patterns without “true/false” labels, which means low‑frequency facts are hard to get right. OpenAI argues calibration can be cheap, and that smaller models may be better at knowing their limits.

Why it matters: Metrics drive roadmaps, procurement and PR. If leaderboards start rewarding “I don’t know,” we could see safer systems and better trust, but also more abstentions and fewer direct answers. Buyers should ask how a model balances accuracy, abstention and error costs, since those trade‑offs shape real‑world value.

In Brief

• South Korea rethinking US investments after ICE detains legal temporary factory construction workers

• Google admits that the open web is in “rapid decline”

Market Trends

Anthropic endorsing California bill to regulate advanced AI
Anthropic has endorsed California’s SB 53, a “trust but verify” bill that would force frontier AI labs to publish safety frameworks, file transparency reports before launch and report serious incidents within 15 days. They argue this codifies practices they and rivals already follow, which mean disclosure becomes mandatory rather than a nice-to-have. They still prefer a federal approach, but say AI progress will not wait for Washington.

• What SB 53 would require. Companies training the most powerful models must “develop and publish safety frameworks,” release pre-deployment transparency reports, report critical incidents, and protect whistleblowers, with penalties for breaking their own commitments.

• Who is covered and who is not. The bill targets large developers above a 10^26 FLOPS training threshold and grants exemptions for startups and smaller firms. Anthropic says this focuses on real risk, although critics may note it largely formalises the incumbents’ playbook.

• Open questions and next steps. Anthropic backs stronger detail on tests and mitigations, and wants regulators able to update rules as capabilities shift. They warn the threshold could miss some models, which suggests calibration will be a moving target.

Why it matters: California often sets the template for tech policy. If SB 53 passes, disclosure-first safety could become the de facto standard for frontier AI, shaping how labs scale and how investors judge risk. The flip side is that rules written around today’s giants may entrench them, so leaders should watch where thresholds land and how enforcement actually bites.

Anthropic agrees to $1.5 billion settlement in copyright case
Anthropic has agreed to a reported $1.5 billion settlement with authors who alleged the startup trained on pirated books, which includes roughly $3,000 per title plus interest and the destruction of datasets, according to CNBC citing a court filing. If approved, it would be the largest publicly reported copyright recovery on record. The deal arrives weeks after a judge said Anthropic’s training methods were fair use, but allowed a trial on whether the company used titles from sites like Library Genesis.

• Terms and scale. The settlement would compensate authors and require Anthropic to delete datasets containing the disputed works. Plaintiffs’ counsel said it “sends a powerful message” to AI firms scraping pirate sites.

• Legal backdrop. In June, the court found training as such to be fair use, yet kept claims tied to allegedly pirated sources alive. Anthropic’s deputy GC said the agreement would resolve “remaining legacy claims” and reiterated its commitment to “safe AI systems.”

• Signals for the industry. Some legal voices called recent rulings a “green light” for common training approaches, while warning “it’s not an all clear.” High‑profile cases like The New York Times v OpenAI and Disney v Midjourney still loom.

Why it matters: For AI leaders, the headline is simple. Training on public corpora may pass fair use, but provenance matters. The cost of tainted data can include billion‑dollar cheques and forced dataset deletions, which mean data governance, licensing and retention controls are now strategic, not optional.

OpenAI expects to burn $115 billion through 2019
OpenAI now expects to burn $115 billion in cash through 2029 as compute and data centre costs surge, even while it lifts revenue forecasts to roughly $200 billion in 2030. Investors are still piling in at a $500 billion valuation, betting the company can turn ChatGPT’s growth and new monetisation schemes into profits. The ramp explains Sam Altman’s line about OpenAI being the “most capital intensive” startup.

• Owning more of the stack. To curb cloud rent, OpenAI plans nearly $100 billion on servers it controls and is pursuing custom chips, which could let it rent capacity to others, similar to AWS, if it ever gets public-market access to cheaper capital.

• Revenue hopes and big assumptions. 2024 revenue is projected at $13 billion, with ChatGPT at nearly $10 billion this year and nearly $90 billion by 2030. They also model $110 billion from “free user” monetisation and Facebook-like margins, although the path is vague and Microsoft takes 20% of revenue.

Why it matters: AI is becoming an infrastructure business which means capex and balance sheets decide winners. Leaders should expect higher model prices, tougher vendor lock-in and more vertical integration, while asking the dull but vital question: do the monetisation assumptions really cover the compute bill.

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.