Weekly Intel - 2026-04-26

The theme this week is succession, in every sense. Apple is handing off leadership, Google is betting its chip roadmap on agents that don’t fully exist yet, and AI-generated music is quietly displacing human artists on streaming platforms. The question running through all of it: when the old guard steps aside, is what’s replacing it actually ready?

Tech Industry

John Ternus to become Apple CEO Apple is making its second CEO transition ever, moving Tim Cook to executive chairman and elevating hardware engineering chief John Ternus to CEO effective September 1. The choice of Ternus over other senior leaders is a clear signal: Apple’s next chapter is defined by physical products (headsets, cars, devices), not services or software alone. Cook staying on as executive chairman with a policy engagement role keeps Apple’s geopolitical relationships intact, and what I’ll be watching is whether a hardware leader can maintain the margins and ecosystem lock-in that made Apple a $3 trillion company.

Tim Cook’s Impeccable Timing Tim Cook’s announcement that he’ll move to Executive Chairman on September 1 is the occasion for a proper CEO eulogy, and the numbers are staggering: 303% revenue growth, 354% profit growth, and a 1,251% increase in Apple’s market value over 15 years. What makes Cook’s tenure genuinely remarkable is that it began under the worst possible narrative conditions: stepping in for a dying Steve Jobs. And yet he turned operational excellence and financial discipline into a $4 trillion company. The question now is whether Apple’s next CEO inherits a machine that runs itself or a strategy that was always inseparable from Cook’s particular genius for timing.

AI Industry Moves

Anthropic takes $5B from Amazon and pledges $100B in cloud spending in return Amazon poured another $5 billion into Anthropic, bringing its total to $13 billion, but the more important number is the $100 billion in AWS spending Anthropic committed to over the next decade, locked to Amazon’s custom Trainium chips through at least the fourth generation. This is the same playbook Amazon ran with OpenAI two months ago: structure the investment so the cash flows right back as cloud revenue. What’s emerging is a model where Big Tech doesn’t just fund AI companies. It becomes their landlord, their chip supplier, and their infrastructure layer all at once. I keep thinking about how much strategic flexibility AI companies actually retain as these dependency loops tighten, and what that means for enterprises betting on them as independent vendors.

OpenAI ad partner now selling ChatGPT ad placements based on “prompt relevance” StackAdapt is pitching advertisers on a limited pilot to run ads inside ChatGPT, with CPMs starting at $15 and targeting based on what users are actively asking about, what they’re calling “prompt relevance.” This is the commercialization of intent at its most raw: not inferred from browsing behavior or search keywords, but extracted from natural language conversations where people are genuinely thinking through decisions. Whether this becomes the highest-signal channel in digital marketing or gradually erodes the trust that makes ChatGPT useful is probably the most interesting tension in AI monetization right now. I don’t think it has to be one or the other, but I haven’t seen anyone argue convincingly that it won’t be.

Infrastructure

Our eighth generation TPUs: two chips for the agentic era Google announced two new TPU chips: one optimized for inference (TPU 8i) and one for training (Virgo), both purpose-built around the specific demands of reasoning-heavy, agentic AI models like Gemini. The notable move is the full-stack co-design: custom silicon, custom ARM-based CPU hosts, custom network fabric, all tuned as a single system rather than assembled from generic parts. This is Google making a vertical integration bet similar to what Apple did with mobile. Whether cloud customers will trade flexibility for that kind of tightly coupled performance is still an open question, though Google is careful to keep open framework support (PyTorch, JAX, vLLM) on the table to soften the lock-in feel.

AI & Software

Deezer says 44% of songs uploaded to its platform daily are AI-generated Nearly half of all new music hitting Deezer’s platform is now AI-generated: 75,000 tracks a day. Yet AI accounts for just 1-3% of actual listening, and 85% of even that is fraudulent. This is a supply-side flood, not a demand story. Deezer is catching most of the fake streams, so the immediate revenue loss is manageable. What’s less manageable is the infrastructure cost of ingesting, analyzing, and policing millions of junk uploads every month, a tax that every platform marketplace (music, apps, e-commerce) is about to pay as generative AI makes production nearly free. When production cost hits zero but filtering cost doesn’t, the bill lands somewhere. Right now it looks like it’s landing on the platforms.

OpenAI releases GPT-5.5 and GPT-5.5 Pro in the API OpenAI dropped GPT-5.5 into its API with a 1M token context window, native computer use, a hosted shell, and a Pro tier that throws more compute at harder problems. The feature list reads less like a language model and more like a general-purpose digital worker: it can browse the web, use tools, patch code, and operate a computer, all within a single API call. For any company building on OpenAI’s stack, the thing worth thinking through now is how fast this collapses the gap between “AI assistant” and “AI agent that actually does the work,” and whether your current automation roadmap still makes sense if that gap closes in 12 months rather than 36.

DeepSeek v4 DeepSeek dropped V4 with open weights: a 1.6 trillion parameter model (49B active) claiming parity with the best closed-source models, plus a lean 284B/13B flash variant, both supporting 1M token context. The active parameter counts matter because they translate directly to inference cost. DeepSeek is pushing hard on the idea that frontier-level performance no longer requires frontier-level spend. What I’m watching is the compounding effect: every release like this compresses the timeline between “only OpenAI or Google can do this” and “anyone can run this on rented GPUs.” If your AI strategy depends on one provider’s moat staying intact, the erosion is moving faster than most planning cycles account for.

Cybersecurity

Bitwarden CLI compromised in ongoing Checkmarx supply chain campaign Bitwarden, one of the top three password managers by enterprise adoption and used by over 50,000 businesses, had its CLI build compromised through a poisoned GitHub Action in its CI/CD pipeline. This is part of a broader Checkmarx supply chain campaign hitting multiple repositories with the same vector: not a vulnerability in the product itself, but in the automated infrastructure that builds and ships it. Attackers are increasingly targeting the toolchain rather than the code, which means your security posture is only as strong as every dependency in your build pipeline. If your organization uses Bitwarden CLI or any tool built with GitHub Actions, start with this: do you have visibility into your build pipeline dependencies at all?

Investigation uncovers two sophisticated telecom surveillance campaigns Citizen Lab identified two separate campaigns where surveillance vendors posed as legitimate cellular providers to exploit SS7 weaknesses and track phone locations, with researchers saying this is likely just a fraction of a much larger problem. SS7 vulnerabilities have been known for years, and the global telecom infrastructure still hasn’t been meaningfully hardened against them. If your executives travel internationally or handle sensitive negotiations over mobile networks, the operational assumption should be that location tracking is already happening. The only useful question from there is whether your security posture reflects that.

That’s what I’m watching. What caught your attention this week?

-Eric