Weekly Intel - 2026-06-07

A pattern ran through most of what I read this week: enormous piles of money are chasing AI and energy infrastructure, and the institutions that decide who’s officially arrived keep moving the line.

AI Industry Moves

Anthropic confidentially submits draft S-1 to the SEC Anthropic filed a confidential S-1, which means the most safety-forward AI lab is now on a path to going public. The filing is thin by design (no share count, no price, no timeline), but the intent reads clearly enough: Anthropic wants the option to tap public markets, almost certainly to fund the brutal capital requirements of building frontier models. This is the step that turns AI from a venture-backed race into a public-market one.

Alphabet announces $80B equity capital raise to expand AI infra and compute Alphabet is raising $80 billion in equity, not debt, to fund AI infrastructure and compute. For a company that throws off cash the way Google does, that’s a telling choice. It says the scale of investment Google thinks it needs to stay competitive now runs past what even its balance sheet can comfortably absorb alone. I keep seeing the same logic across Big Tech: the cost of not spending gets treated as existential, so the spending never stops, and this raise makes that math explicit. If you build on or compete with Google’s cloud and AI platforms, the thing I’d be modeling is this: when the infrastructure layer gets capitalized at this size, what does it imply about the pricing power, dependency, and switching costs you’ll be living with in 18 months?

OpenAI frontier models and Codex are now available on AWS This one is about plumbing more than technology. OpenAI’s models and its Codex coding agent are now generally available through Amazon Bedrock, so enterprises can reach them inside their existing AWS security, compliance, billing, and governance setup, with no new vendor to onboard. To me the signal is that the model layer is commoditizing into infrastructure fast. What matters more than which model you pick is how fast your organization moves from pilot to production, and dropping these models into AWS compresses that gap for the millions of companies already running there.

Financial Markets

S&P 500 rejects SpaceX, also blocking entry for OpenAI and Anthropic The S&P refused to bend its eligibility rules for SpaceX, and that decision most likely shut the door on OpenAI and Anthropic, too. No exceptions on the profitability screen, the minimum public float, or the standard one-year seasoning period. SpaceX alone would have triggered something like $14 billion in passive fund buying on a fast-track inclusion, with OpenAI and Anthropic pulling in billions more. Strip away the index-methodology talk and what’s left is plain: the most capital-hungry AI companies are still structurally unprofitable, carrying heavy debt to fund their buildouts, and the one mechanism that would have handed them a wave of automatic institutional money just said no. So where does the next round of patient capital actually come from?

AI & Software

When AI Builds Itself: Our progress toward recursive self-improvement Anthropic published something unusually candid: internal data showing its engineers now ship 8x more code per quarter than they did between 2021 and 2025, mostly because AI is doing more of the work. The piece sketches a path toward “recursive self-improvement,” AI systems designing their own successors, and argues it could show up before most institutions have any plan for it. Anthropic isn’t publishing this as a theoretical exercise. It reads like a company telling you the feedback loop is already speeding up. If you run a technology-dependent organization, that’s worth taking seriously now, because a strategy built around human engineering timelines may not survive a capability curve that no longer waits on them.

Gemma 4 12B: A unified, encoder-free multimodal model Google’s new Gemma 4 12B is multimodal (text, vision, native audio) and runs on a laptop with 16GB of VRAM, no specialized encoders required. Vision and audio feed straight into the language model backbone, which simplifies the architecture and gets benchmark numbers close to Google’s larger 26B model. Capable AI keeps moving off the cloud and onto local hardware, which changes the math on data privacy, latency, and infrastructure cost. At some point “runs on a laptop” becomes the baseline expectation for enterprise AI tooling, and a lot of cloud-dependent workflows are going to look expensive by comparison.

Artificial intelligence is not conscious – Ted Chiang Ted Chiang goes straight at Anthropic’s increasingly human framing of Claude: the 84-page “constitution” written with Claude as its “primary audience,” the CEO’s openness to AI consciousness, the careful language about “functional emotions.” His argument is that this is marketing dressed up as uncertainty, and that it makes it harder for everyone to think clearly about what these systems actually are. If you’re making deployment, governance, or investment calls around AI, Chiang raises a sharp question: do the companies building these tools have an incentive to blur the line between performance and sentience, and what does that blur cost you when they do?

Energy & Transportation

Wind and solar generated more power than gas globally in April 2026 For the first time, wind and solar outproduced gas worldwide for a full month: 531 TWh against 477 TWh in April. The crossover is the headline, but the detail that got my attention is that gas output barely budged in five years while renewables more than doubled. That gap, widening against the backdrop of yet another Middle East energy scare, is the clearest sign yet that any firm still anchoring its long-term planning to gas-price assumptions is modeling a world it has already fallen behind.

Tech Industry

Google will pay SpaceX $920M per month for compute Between the Google and Anthropic deals alone, SpaceX is now locking in roughly $2.2 billion a month in GPU rental revenue, all ahead of its IPO. That’s a company building a recurring-revenue infrastructure business the size of a hyperscaler’s while also being the hyperscalers’ landlord. Google runs one of the largest cloud platforms on the planet and is still paying a rocket company nearly $1 billion a month for GPU access it apparently can’t stand up fast enough on its own. If the biggest cloud providers are now net buyers of someone else’s compute, where do the real bottlenecks and the real leverage actually sit in the AI value chain?

Meta launches Instagram, Facebook, and WhatsApp subscriptions Meta is now selling consumer subscriptions across Instagram, Facebook, and WhatsApp worldwide: a few dollars a month for profile customization, super reactions, and story insights, while it tests professional tiers for creators, businesses, and AI users under one “Meta One” brand. The consumer features look like table stakes. The infrastructure underneath is the actual play. Meta is building a direct billing relationship with billions of people, and that changes the math on everything it does next, from AI distribution to commerce to its dependence on advertising. Once that layer is running across three of the most-used apps in the world, what Meta bundles into it next is the part I’d keep an eye on.

That’s what I’m watching. The S&P ruling is the one I’ll be tracking. Rules like that rarely bend on the first push, but they don’t always survive the second.

-Eric