Daily Digest — June 12, 2026

[BUSINESS] SpaceX Goes Public — Biggest IPO Ever, at 100x Revenue

The story: SpaceX priced its IPO at $135/share, raising $75 billion at a $1.77 trillion valuation. The offering was 4x oversubscribed, with 20% of shares reserved for retail investors who placed $100 billion in buy orders. An estimated 4,400 employees become paper millionaires today. Trading begins on Nasdaq under SPCX.

My take: The valuation is roughly 100x revenue, which doesn’t work by any traditional metric. What investors are actually buying is the vertical integration story — reusable rockets are the unlock for orbital datacenters, which are the unlock for being a monopoly provider of marginal compute capacity. Nobody else can put chips in space cheaply because nobody else has the launch infrastructure. But the part that’s underdiscussed is maintenance. Terrestrial datacenters have humans who swap failed drives, replace cooling units, and patch firmware. In orbit, satellites are disposable — if something breaks, it becomes space junk. The economics only work if failure rates stay low enough to depreciate and replace faster than things break. That’s an engineering problem nobody has solved at scale yet, and it’s the biggest variable in whether the orbital datacenter thesis holds.

[BUSINESS] Bezos Bets on Manufacturing AI — Prometheus Raises $12B

Source: Morning Brew · TLDR

The story: Jeff Bezos’ startup Prometheus raised $12B at a $41B valuation, seven months after launching with ~$6B from Bezos himself. The company is building an “artificial general engineer” — an AI system trained on physics laws and manufacturing data to accelerate the design and production of physical products. Bezos is reportedly looking to raise $100B to acquire industrial companies that could feed data to the model.

My take: Everyone in AI is focused on software — code generation, chatbots, agents. Bezos is betting on atoms, not bits, and he’s one of the few people with credibility there. Amazon’s fulfillment centers are among the most efficient physical operations on the planet. He’s seen what optimization at scale looks like in the real world and is now trying to apply AI to the design loop itself. The $100B acquisition target is the tell — this isn’t a software play, it’s a data moat play. Whoever has the best physics and manufacturing data trains the best model, and you can’t distill that out of a public API. That’s a defensible position in a way that most AI companies aren’t.

[AI] Fable’s Backlash Is Real — Codex Usage Spiked as Customers Walked

Source: The Pragmatic Engineer · Gergely Orosz

The story: Anthropic’s Fable 5 launch backfired commercially. Codex’s Head confirmed a strong spike in token consumption immediately after Fable launched — “unusual when we don’t launch something.” Microsoft told employees Fable is off-limits while legal evaluates the 30-day data retention requirement. AWS confirmed Fable data leaves their security boundary even on Bedrock. Cursor disabled Fable in privacy mode. After intense backlash, Anthropic partially walked back the silent nerfing — they’ll now make model degradation visible — but didn’t budge on data retention.

My take: The AI disaster recovery analogy is the right frame here. Enterprises already run DR exercises for databases, cloud providers, and networking. Now they need the same thing for AI models — can we fail over from Anthropic to OpenAI in 24 hours without breaking production workflows? Instead of testing cross-zone failover on a weekend, you’re routing traffic to a different model provider to prove you can. The companies that treat model dependency the same way they treat infrastructure dependency are going to be better positioned. The ones going all-in on a single provider with no exit plan are building the same kind of fragility that Coinbase just demonstrated with its single-AZ AWS dependency — a 10-hour outage because there was no automated failover.

[ENG] Smart Model Routing Is the Next Infrastructure Layer — And AI Gateway Should Be in the List

Source: The Pragmatic Engineer · Gergely Orosz

The story: Pragmatic Engineer catalogued the emerging smart model routing space — tools that automatically route queries to the cheapest or most capable model, cutting token costs by 20-30%. Token prices between models can differ by 10-20x, so routing “easy” requests to cheap models and “hard” ones to frontier models is an obvious lever. The list includes Factory Router, Not Diamond, OpenRouter, Weave, LiteLLM, and several AI gateways. Cloudflare AI Gateway wasn’t mentioned.

My take: Cloudflare AI Gateway already has the infrastructure to do this — it sits in front of model calls, logs cost per request, and can route by provider. Adding role-based or complexity-based routing rules is a natural extension of what already exists. The fact that it wasn’t mentioned in the Pragmatic Engineer’s list means either the product team hasn’t positioned it this way or the developer community doesn’t know it can do this. Either way, that’s a gap. The smart routing market is going to be significant because enterprise AI budgets are ballooning and CFOs want the same cost visibility and optimization they get from tools like Vantage or CloudHealth. The gateway layer is exactly where that belongs.

[ENG] Cloudflare Scales Security Insights 10x — What It Actually Means for Customers

Source: Cloudflare Blog · Dave Baxter

The story: Cloudflare’s Security Insights team scaled scanning throughput from 10 scans/second to 120+, through engineering improvements rather than throwing more hardware at it — parallel Kafka consumption, fast/slow lane splitting, bulk database writes, and fixing a cross-continent latency bug where Amsterdam API calls were hitting a Portland database and taking 3 seconds instead of 10ms. Practically this means: automatic scanning now enabled for all free accounts, scan frequencies doubled — free accounts weekly, Pro/Business every 3 days, Enterprise daily.

My take: For enterprise customers the actionable output is daily scans surfacing specific security misconfigurations — dangling DNS records, missing security headers, exposed origins, misconfigured SSL settings — with one-click remediation guidance from the Security Overview dashboard. That’s not a passive feature, that’s something worth bringing up in customer conversations. The engineering story is also a useful internal reference point — they fixed a 300% latency problem that was hiding in plain sight because metrics showed “API timeouts” rather than “your Amsterdam instance is talking to Portland.” The lesson that you can 12x throughput by fixing your code before adding capacity is worth bookmarking.