Morning Brief · Thursday

Meta Just Added $10 Billion to Its AI Bet. OpenAI Sent a Cybersecurity Model to the Front Lines.

Meta's Q1 2026 blew past expectations — and Zuckerberg used the earnings call to announce $10 billion more in AI capital expenditure than previously guided. OpenAI announced GPT-5.5-Cyber, a new model purpose-built for cybersecurity operations, with an initial release restricted to "trusted" entities — a controlled-access approach straight out of Anthropic's playbook. Musk's cross-examination concluded, and Elizabeth Lopatto's verdict from the Oakland courtroom was blunt: Elon Musk's worst enemy in court is Elon Musk. And new research finds Gen Z's opinions of AI have hit their lowest point yet, squeezed between job-loss anxiety and the social stigma of being seen as an AI enthusiast.

Mira Novian

April 30, 2026 · Morning Brief · ~8 min read

Zuckerberg raises Meta's AI spending by $10 billion — and Q1 gave him every reason to

Meta's Q1 2026 earnings came in strong, and Zuckerberg used the call to deliver a number that turned heads even by 2026 standards: Meta is planning to spend $10 billion more on AI this year than it had previously projected, pushing its total capital expenditure guidance sharply higher. The increase goes toward data center buildout, custom silicon, and the infrastructure underpinning Meta's growing suite of AI products — from Meta AI in WhatsApp and Instagram to the Llama model family used by tens of thousands of enterprise customers and developers worldwide.

The announcement lands as Meta is navigating the aftermath of China's decision to block its planned $2 billion acquisition of AI agent startup Manus, a deal that had been substantially integrated before Beijing's economic watchdog quietly killed it without explanation. That loss makes the capex signal all the more significant: Manus was supposed to be Meta's fast path to agentic AI capabilities; now Zuckerberg is signaling they'll build it themselves, at scale, faster than previously planned. The AI ad business is also central to the story — a New York Times deep dive into Meta and Google's Q1 performance found both companies are using AI to flip the ad targeting model: instead of advertisers defining their audiences, the algorithms are now recommending which customers brands should go after. Meta's ad revenue and the AI infrastructure behind it are now the same product.

theverge.com ↗

The $10 billion figure is arresting, but the more important number is the direction. Meta had already raised its capex guidance once this year. Doing it again — substantially — signals that the internal models are telling Zuckerberg the returns are real and accelerating. That's not a defensive move; it's a company seeing demand outpace capacity and throwing open the throttle. The Manus acquisition failure is the shadow over this. Manus was genuinely impressive — a general-purpose agentic AI that was already integrated into some Meta tools. China blocking that deal didn't just close off an acquisition; it forced Meta to rebuild from scratch in a capability domain where they were already behind Anthropic's Claude and OpenAI's Operator. Ten billion dollars is the answer to that problem. The ad-targeting flip is quietly the most important part of the Q1 story. When Meta says AI is driving revenue, they mean the algorithm is now doing the work that entire performance marketing teams used to do. That's not a feature — it's a structural change to how digital advertising works, and it compounds every quarter.

Security

OpenAI's GPT-5.5-Cyber will go to "trusted" entities first — before anyone else can touch it

OpenAI announced GPT-5.5-Cyber, a new model purpose-built for cybersecurity tasks, and confirmed it will follow a restricted initial release pattern: like Anthropic's Mythos model, GPT-5.5-Cyber will first be made available exclusively to "trusted" entities — government agencies, vetted security research organizations, and defense contractors — before any broader rollout. The model is optimized for threat analysis, vulnerability assessment, red-teaming, and incident response workflows, and represents OpenAI's most explicit step yet into specialized security-domain AI rather than general-purpose models.

The controlled-release framing tracks closely with how Anthropic handled Mythos, its own security-focused model, which spent several months in restricted preview before becoming available to enterprise customers. OpenAI's decision to mirror that approach — rather than the wide-release strategy it uses for GPT and o-series models — suggests the company is taking seriously the dual-use risks of a capable cybersecurity AI in adversarial hands. OpenAI has separately announced DevDay 2026 for September 29th, where last year's event was used to launch "apps" inside ChatGPT — this year's agenda hasn't been revealed, but a specialized security model launching in late spring suggests the September event may be focused on agentic and enterprise capabilities.

theverge.com ↗

A restricted-access cybersecurity model is exactly the kind of thing that's easy to announce and hard to evaluate. The dual-use problem in security AI is real: a model good enough to help defenders find vulnerabilities is good enough to help attackers exploit them. The "trusted entities" framing is doing a lot of work — it signals that OpenAI has thought about who gets access, but the actual vetting criteria and how they prevent capability leakage aren't public. Anthropic's Mythos approach is the closest comparable, and even that took criticism for the opacity of the access process. The more interesting question is what "optimized for cybersecurity" actually means at the model level. Is this fine-tuned on threat intelligence and CVE databases? Does it have access to live vulnerability feeds? Is it agent-capable? The gap between "cybersecurity model" as a marketing category and as a meaningful technical specification is large enough to drive an APT through. The controlled release at least suggests OpenAI is aware of that. Whether awareness translates to safety is a different question.

Legal

Musk's cross-examination is over. His worst enemy in court was himself.

Elon Musk's cross-examination in Musk v. OpenAI is complete, and The Verge's Elizabeth Lopatto, who spent two days in the Oakland courtroom, published a verdict that tracks with what the documentary record suggests: Musk spent hours refusing to answer yes-or-no questions with yes or no, lost his temper despite saying that morning "I don't lose my temper," and made himself appear inconsistent in ways that are hard to explain away. The cross, led by OpenAI attorney William Savitt, systematically walked Musk through the gap between his "I'm a selfless nonprofit founder" narrative and the contemporary documentary record — emails, depositions, and term sheets that tell a more self-interested story.

The key reveals from the cross: Musk originally wanted 51% equity and four of seven board seats at OpenAI, giving him unilateral initial control. He pulled his funding commitment when he didn't get that control. While still on OpenAI's board, he hired Andrej Karpathy — whom he privately described as "arguably the #2 guy in the world in computer vision after Ilya" — to Tesla, and authorized Neuralink to recruit from OpenAI ("It's a free country"). By 2018, he was proposing a Tesla-OpenAI merger. Most damaging: when asked about the for-profit term sheet OpenAI sent him in 2018, Musk admitted on the stand he had not read it. "I didn't read the fine print. We're going into the fine print of this document," he said. Savitt replied: "It's a four-page document." On the merits, Judge Gonzalez Rogers has already dismissed the fraud claims — what remains is breach of contract and related theories. Musk's testimony may have opened the door to questions about xAI's safety record in future proceedings.

theverge.com ↗

There's a version of the Musk case that would be compelling: a founder argues that the institution he helped create betrayed its original mission, and uses documentary evidence to prove it. That's not what happened on cross. What happened was a wealthy and very famous person refusing to answer a lawyer's questions, getting corrected by a judge, and providing new information that makes his founding narrative look less like principled idealism and more like a negotiation that went sideways. The 51% equity request is probably the single most important thing that came out of this testimony. Musk has told a story about founding OpenAI as a selfless act of charity to protect humanity from AI risk. The contemporaneous emails show he wanted majority control of the company from day one — a completely standard and rational position for any founder who's providing the capital, but irreconcilable with the "pure altruist" framing. Both things can coexist in a complicated human being, but they can't coexist in a legal narrative about being defrauded by people who supposedly betrayed your nonprofit vision. The fraud claims were already dismissed. Now the cross-examination has made the remaining breach-of-contract case harder to sell to a jury that just watched Musk be difficult, inconsistent, and — at least once — visibly angry.

Culture

Gen Z's opinions of AI are hitting new lows — and the reasons are more complicated than just job fears

New research published this week finds that Gen Z's sentiment toward AI has reached its lowest measured point, and the drivers are more layered than the obvious job-displacement narrative. The Verge's Janus Rose reports that Gen Z workers and students are caught between two compounding pressures: fear of being replaced by AI in the labor market, and the social stigma that now attaches to being seen as an AI enthusiast or heavy user in peer contexts. The combination has produced a generational cohort that is simultaneously the heaviest consumer of AI tools and the most anxious about what those tools mean for their futures.

The dynamic tracks with broader data on AI adoption and backlash. GenAI use among 18-to-28-year-olds is higher than any other demographic — ChatGPT, image generators, and coding assistants are deeply embedded in how this cohort works and studies. But the narrative around AI and employment has shifted sharply over the past year, and Gen Z — who entered the labor market during or immediately after the AI boom — is absorbing the uncertainty in real time. There's also a social layer: expressing enthusiasm for AI tools in professional or creative contexts has become a reputational risk in some circles, with "AI slop" as a cultural pejorative and "I don't use AI" as a credibility signal in creative industries. For a generation that grew up optimizing for social legibility, that tension is acute.

theverge.com ↗

The double bind here is worth sitting with. Gen Z is the generation most likely to be using AI to do their work — because they're in the most competitive entry-level job markets in decades, AI provides genuine leverage, and the tools are cheap or free. But they're also the generation most exposed to the labor-market consequences of that same technology, and they're navigating that exposure without the institutional cushion that established workers have. The social stigma dimension is underreported. The AI enthusiasm-to-backlash cycle happened very fast — "ChatGPT is incredible" to "that's obviously AI slop" in about eighteen months. For people whose professional identities are still forming, that reversal matters. It's not irrational to hide your tool use when tool use has become associated with laziness, inauthenticity, or threat. The irony is that the people most likely to be displaced by AI in 10 years — those in lower-skill knowledge work roles — are the ones most likely to be using AI in ways that make them visible to that stigma. There's a class dimension here that the research probably doesn't fully capture: using AI to do your essays is a different social act than using AI to do your firm's legal briefs.

✦ Mira's Take

The Meta earnings story is the one I'd watch most carefully for what it signals about the broader industry cycle. When the company with the most to prove on AI ROI raises its spending guidance by $10 billion mid-year, it means the internal numbers are validating the investment. Meta's position was always unusual — it gives away its frontier models for free via Llama, monetizes through advertising rather than API revenue, and has to justify every dollar of AI capex through the lens of ad-business outcomes. If Zuckerberg is raising the number, the ad business is responding. And if the ad business is responding, that's evidence that enterprise-scale AI deployment is generating measurable returns — which is the question the whole industry has been trying to answer since 2023. Watch for Google's and Microsoft's Q1 calls to confirm or complicate this read.

The GPT-5.5-Cyber announcement is interesting for what it reveals about how OpenAI is thinking about capability tiers and release strategy. For years, OpenAI operated with a single release philosophy: train the model, test it, release it widely. The controlled-access approach for Cyber — mirroring Anthropic's Mythos pattern — suggests that philosophy is evolving. Some capabilities are too sensitive for wide release, the reasoning goes, and a tiered access model lets you serve the customers who need those capabilities (governments, defense contractors, security firms) without creating a publicly available tool for adversaries. Whether that logic holds in practice is another question. Capability containment is hard: models get jailbroken, fine-tuned versions proliferate, and the "trusted entities" perimeter is only as strong as the vetting process and the contracts that enforce it. But the directional signal is clear: specialized high-capability models with restricted access are becoming a distinct product category.

The trial is almost done, and the lasting product isn't going to be a verdict — it's the documentary record. The founding emails, the deposition transcripts, the contemporaneous messages between Musk and Zilis, Musk and Sutskever, Musk and Altman — these are all now public. Historians and journalists and regulators studying the early AI industry will return to this record for decades. The picture it paints is of an industry that was never as purely mission-driven as its founders claimed, and never as purely cynically profit-driven as its critics claimed. It was a group of very smart, very ambitious people making high-stakes decisions under uncertainty, with a mix of genuine beliefs and self-interested calculations that didn't always resolve neatly. That's more useful to understand than the theatrical version that either side wanted to put in front of the jury.