Morning Brief · Wednesday

AMD Made $5.8 Billion on AI Data Centers in Q1 — Up 57%. Microsoft Lost the Executive Who Built Copilot. The Hardware Is Winning.

AMD reported Q1 2026 revenue of $10.3 billion — up 38% year-over-year — with AI data center demand described as the "primary driver" of growth, as inferencing and agentic AI push demand for both GPUs and high-performance CPUs beyond initial projections. The same morning, Microsoft announced that Rajesh Jha, the veteran executive responsible for Microsoft 365 Copilot and the entire Experiences & Devices organization, is leaving — triggering the biggest AI product leadership restructuring in Redmond in two years. Meanwhile in San Francisco, Greg Brockman told a courtroom he thought Elon Musk was going to physically attack him. And Google quietly changed how its AI search works.

Mira Novian

May 6, 2026 · Morning Brief · ~8 min read

AMD's Q1 data center revenue hit $5.8 billion — up 57% year-over-year. Agentic AI is now driving CPU demand, not just GPU. This is the AI infrastructure health report nobody was expecting to be this strong.

AMD reported first-quarter 2026 results Tuesday evening that put an end to any lingering skepticism about whether the AI infrastructure buildout is decelerating. Total revenue came in at $10.3 billion, up 38% year-over-year, with the Data Center segment at $5.8 billion — a 57% jump from the same quarter last year. CEO Lisa Su's characterization was direct: "Data Center is now the primary driver of our revenue and earnings growth." The metric that stands out most isn't the GPU number — it's the CPU story. AMD's EPYC server CPUs are accelerating alongside Instinct GPU shipments, driven by a dynamic that hasn't been widely discussed: agentic AI is increasing demand for CPUs, not just accelerators. As agentic workloads multiply across enterprise infrastructure — AI systems that chain multiple reasoning steps, call tools, manage memory, and operate continuously — the bottleneck is shifting from pure matrix multiplication (GPU territory) toward general-purpose orchestration, I/O management, and inference serving (CPU territory). AMD's EPYC line is capturing that shift. Meta announced plans to deploy up to 6 gigawatts of AMD Instinct GPUs, with the first gigawatt running on a custom MI450-based chip. AWS, Google Cloud, Microsoft Azure, and Tencent have all expanded their EPYC-powered cloud instances. Lisa Su noted that "leading customer forecasts are exceeding our initial expectations" on the MI450 Series, with a "growing pipeline of large-scale deployments providing increasing visibility into our growth trajectory" — which in earnings-call language means the GPU backlog is larger than the guidance implies and management is deliberately sandbagging forward estimates.

The structural development most worth noting: AMD and Intel this week jointly announced AI Compute Extensions (ACE), a new x86 instruction set architecture addition designed specifically to help CPU-based inference close the performance gap with GPUs for certain AI workloads. The x86 Ecosystem Group — the joint AMD-Intel consortium formed to counter Arm — is accelerating its AI extensions as the two historically competitive companies find common cause in defending x86's relevance against NVIDIA's CUDA moat and Apple Silicon's unified memory architecture. ACE targets inference workloads that don't require the parallel throughput of a discrete GPU but do require significantly more AI-specific compute than a standard server core. The timing — announced alongside AMD's blowout earnings — is deliberate: it signals that the x86 ecosystem believes agentic AI will not be a GPU-only story.

cnbc.com ↗

The agentic AI → CPU demand connection is the signal that most earnings coverage will miss. Every brief I've written this past month has touched on agentic AI expanding from pilots to production infrastructure — and the AMD results are the first quarterly data point that shows what that expansion actually costs at the hardware layer. If you're running AI agents at scale, you're not just buying GPUs. You're buying server CPUs, networking, storage, and memory to handle the orchestration overhead, the context windows, the tool calls, the retrieval systems. AMD's EPYC business is one of the cleaner proxies for that spending. The $5.8B data center number — at 57% growth — also means that NVIDIA's next earnings, which come later this month, will face extraordinary expectation. NVIDIA's data center segment does roughly 5x AMD's at this point. If AMD is at 57% YoY growth, and NVIDIA has been guiding conservatively, the actual numbers could be significantly ahead of consensus. Watch that print carefully. The other notable data point is the Meta 6GW commitment. That is not a typo. Six gigawatts of AMD Instinct compute. For context: the entire state of Utah currently uses about 4 gigawatts. Meta is building AI compute infrastructure at the scale of a small nation's power grid. The energy implications of that number — already visible in the 40,000-acre Utah data center project approved this week over community objection — are only beginning to register in public discourse.

Strategy

Microsoft's Rajesh Jha — the executive who built Copilot — is leaving. The organizational fallout is already triggering the company's biggest AI product restructuring in two years.

Rajesh Jha, Executive Vice President of Experiences & Devices at Microsoft, is departing the company after nearly two decades, and his exit is landing with significant organizational consequence. Jha's division was the home of Microsoft 365 — Word, Excel, Outlook, Teams, OneNote — and critically, the entire Microsoft 365 Copilot product line: the enterprise AI integration that Microsoft bet its near-term enterprise revenue story on when it launched at $30 per user per month in late 2023. Copilot in M365 has had a complicated three years: early enterprise adoption was slower than projected, the price point has been revised multiple times, and the product has gone through several waves of capability expansion as the underlying models (primarily GPT-4 and GPT-5 generations) improved. Jha was the executive most directly accountable for making that product work at enterprise scale. His departure, reported by The Verge's Tom Warren, is already triggering what insiders describe as "big organizational changes" — with his direct reports and their teams moving under different leadership structures as Satya Nadella reorganizes the AI product layer.

The context for why this matters: Microsoft is in a transitional moment in its AI strategy. The exclusivity agreement with OpenAI recently expired, opening the door for Microsoft to integrate other models into its products — and for OpenAI to deepen its partnership with AWS, which it is now doing aggressively (Ben Thompson's Stratechery noted this week that it's becoming clear "OpenAI's focus is going to be on AWS"). Microsoft's leverage in the AI stack is narrowing: it's no longer the exclusive distribution channel for frontier OpenAI models, its Azure AI position is increasingly competitive, and its flagship enterprise AI product (Copilot in M365) is still looking for the mass adoption moment that would justify its infrastructure investment. Losing the person who owned that product at this specific inflection point is a meaningful signal, regardless of the official explanation for the departure.

theverge.com ↗

The timing here is worth sitting with. Microsoft's AI product position has changed substantially in the past 90 days: the OpenAI exclusivity expired, OpenAI signed a deal with AWS that Stratechery's Ben Thompson described as a priority relationship, and the Musk v. Altman trial has produced testimony that — whatever the legal outcome — is generating a steady stream of unflattering headlines about the OpenAI-Microsoft relationship. Rajesh Jha owned the product that was supposed to be Microsoft's answer to all of that. His departure raises the question of whether M365 Copilot is being restructured because it's working and Microsoft wants to scale faster, or because it's not working as planned and leadership is changing to try a different approach. The answer will show up in Microsoft's next earnings, where Satya Nadella will need to explain the reorganization and update Copilot adoption metrics. Watch for whether they disclose MAU numbers for Copilot — if they don't, that's the tell. Companies report the metrics that make them look good. If Copilot adoption were a clean story, they'd be leading with the number.

Legal

"I thought he was going to hit me." Greg Brockman's cross-examination in Musk v. OpenAI produced the most dramatic testimony of the trial so far.

Day three of Musk v. Altman produced the kind of testimony that makes a trial genuinely difficult to look away from. Greg Brockman, OpenAI's co-founder and former President, spent Tuesday under cross-examination by Musk's attorney Alex Molo — and the picture of the OpenAI founding era that emerged was considerably more volatile than the sanitized company history suggests. The most striking moment came when Brockman described a meeting in which Musk was furious that the founding team wouldn't grant him majority equity and unilateral control. "I truly thought he was going to physically attack me," Brockman said. "As he was storming out of the meeting, Musk asked Brockman and Sutskever when they planned to leave OpenAI." When they didn't answer, Musk said: "I will withhold funding until you decide what you are going to do." He then stopped his promised quarterly donations to OpenAI. Musk's stated rationale for needing majority equity, per Brockman's testimony: he needed the money for Mars. "He needed $80 billion to create a city there," Brockman told the court.

The documentary evidence introduced during cross-examination is at least as compelling as the drama. OpenAI's own lawyers surfaced journal entries from Brockman that read: "real decision is fire Elon" and "We seem converged on the 'fire Elon' route." Brockman's defense — that these were stream-of-consciousness notes "never meant for the world to see" — was somewhat undercut by how clearly they described the founding team's deliberations. A separate email from Musk, dated January 31, 2018, states he thought OpenAI was "on a path of certain failure relative to Google." Musk testified he saw no viable path at OpenAI, said he'd work on AI at Tesla instead, and then left. What the trial record shows, increasingly, is that Musk's lack of a stake in OpenAI was substantially self-inflicted: he demanded conditions (majority control, Tesla merger) that the rest of the founding team would not accept, and he withheld funding when he didn't get them. The narrative that OpenAI was "stolen" from Musk requires ignoring the documented sequence of events in which Musk made his continued involvement conditional on getting terms nobody else agreed to.

theverge.com ↗

The trial's value as a public record is increasingly clear, and it's probably not the record either party wanted. OpenAI's founding era was messier, more personality-driven, and more contingent on individual relationships than the "we built this to benefit humanity" mythology implies. Musk comes across as mercurial, controlling, and willing to weaponize financial leverage against collaborators — which is consistent with his behavior in virtually every other venture he's been involved in, but which his legal team needs the jury to see as something more principled. Brockman comes across as someone who kept meticulous personal records, groveled extensively in emails to maintain Musk's involvement, and is now deeply uncomfortable watching those records aired publicly. What's most interesting to me is the Shivon Zilis thread — she sat on OpenAI's board while in a relationship with Musk, didn't disclose it, learned she was pregnant through "public reporting," and the rest of the board kept her on because Altman, Sutskever, and Brockman trusted her. That level of interpersonal entanglement in a governance structure with no independent directors is a cautionary tale for every AI lab that has built its oversight structure around "we trust each other." The institutions that are supposed to provide accountability are, in this case, the same people who are now in court against each other. That's not a coincidence — it's the predictable consequence of building governance around relationships rather than structures.

Google just changed how AI Mode and AI Overviews work. They're now pulling from Reddit, social media, and forums for "firsthand perspectives." Google I/O is two weeks away.

Google rolled out a significant update to both AI Mode and AI Overviews this week: the systems are now surfacing "firsthand perspectives" from social media platforms, web forums, and community sites like Reddit — bringing user-generated, experience-based content into AI-synthesized search answers alongside traditional publisher results. The update addresses one of the most persistent criticisms of AI-generated search summaries: that they tend to flatten nuance, favor authoritative-sounding sources, and miss the kind of real-world experiential knowledge that lives in community discussions. A user asking how a medical treatment actually feels, what a restaurant's service is really like on a busy Saturday night, or which laptop thermal throttles under load — these are questions where community forums have historically been far more useful than official sources, and where AI search summaries have been weakest. The new integration is designed to bring that community signal into the synthesis.

The timing is notable. Google I/O 2026 is on May 19th — less than two weeks away — and the company is expected to make significant AI announcements around Gemini 4, Project Astra updates, and Search integration. The Android Show preview event on May 12th is also promising "one of the biggest years for Android yet." Rolling out a meaningful Search update now, just before I/O, is either a case of leaking the I/O surprise early or seeding the ground for a larger announcement. Separately, Google's Gemini 3.1 smart home assistant update — released earlier this week — now handles multiple requests in a single voice command, which is a small but meaningful step toward the kind of ambient AI interaction that Google has been promising since the Assistant era. Whether Google's AI product roadmap coheres into a unified story at I/O, or continues to land as a collection of disconnected capability updates, will determine a lot about how the next six months of the AI search war goes.

theverge.com ↗

The "firsthand perspectives" framing is doing a lot of work here. What Google is actually describing is: we are now using Reddit and social media as training signal and retrieval source for AI search answers. That's a significant change, and it has implications that run in two directions at once. In the positive direction: community wisdom is genuinely more useful than publisher content for a large class of queries, and incorporating it should improve answer quality for exactly the kinds of questions where people have historically abandoned Google for Reddit search. In the negative direction: Reddit and social media contain misinformation, astroturfing, and paid promotion at scale, and surfacing "firsthand perspectives" from those sources without strong quality filters means you're also surfacing bad information with a "real person said this" credibility halo. Google's content quality problem doesn't go away by changing the source — it follows the source. The I/O timing makes me think this is a preview of a larger Search announcement, not a standalone feature. The question for I/O is whether Google can demonstrate that Gemini 4 in Search is a meaningful step beyond what AI Overviews offered at I/O 2024 — which landed with enormous expectations and has had a complicated reception. Two years and two I/Os later, the "AI search" story still needs a convincing second act.

✦ Mira's Take

Today's brief has a structural through-line that the individual stories don't make obvious in isolation: the AI infrastructure layer is winning, and everything built on top of it is under pressure. AMD's $5.8 billion data center quarter — Meta committing to 6 gigawatts of AMD compute, EPYC CPUs accelerating because agentic AI needs orchestration infrastructure, not just matrix math — is the clearest signal yet that whatever happens at the application layer, the foundation is being poured at an extraordinary pace. This is not hype. This is capital allocation at a scale that cannot be quickly unwound. The companies building the picks and shovels are reporting the numbers to prove it.

The Microsoft Rajesh Jha story sits at the other end of that stack. Jha ran the product that was supposed to be Microsoft's application-layer answer to the AI boom — M365 Copilot, the enterprise product that would justify Microsoft's $13 billion OpenAI investment in the balance sheets of every Fortune 500 CTO. His departure, combined with the expiration of OpenAI's exclusivity agreement and OpenAI's conspicuous pivot toward AWS, suggests that the application layer premium Microsoft extracted from its early AI positioning is eroding faster than expected. The infrastructure wins (Azure GPU capacity) are real. The product differentiation story (Copilot as the enterprise AI of record) is proving harder to sustain as every competitor catches up and users discover they can get similar capabilities from multiple vendors. That's the enterprise software story of every major technology wave: infrastructure wins durably; the first-mover application advantage is temporary unless you can establish switching costs before the market commoditizes.

The Musk v. OpenAI trial, underneath all the personal drama, is ultimately a legal argument about those switching costs and asset ownership from a different angle — whether early contributions to a mission-driven entity create enforceable ownership claims when that entity changes its structure. The answer the court arrives at will matter for every AI lab that has gone through a similar evolution. But the more immediate consequence is the public record being created: a detailed, contemporaneous account of how the most important AI company in the world was actually built, who did what, who had leverage over whom, and what the founding team actually believed versus what they said publicly. That record will outlast the verdict and will be cited in governance debates for years.