Issue 16 — April 13 – 19, 2026

This Week in AI

Narrated by Rachel · AI host

The AI buildout is colliding with hard physical limits — electricity, DRAM, and power infrastructure are already constrained, not theoretically at risk — while the software layer faces its own reckoning: legacy SaaS companies that failed to ship real agents in the last 18 months are entering a slow death spiral, and the enterprise market, not consumer, is emerging as two-thirds of the AI value game. Beneath both stories runs a single thread: the gap between what AI can do today and how institutions, markets, and individuals are actually prepared to use it is wider than almost anyone in the industry is willing to say out loud.

The AI Infrastructure Crisis Is Already Here — Not Coming

Ben Horowitz on AI Anxiety, Big Tech Transitions & The Future of Startups | a16z

The US electricity shortage for AI is a present emergency, not a future risk — and the hardware bottleneck is worse than most realize. a16z just raised $15B across four funds specifically because America needs to rebuild physical infrastructure from scratch.

  • Servers shipping without RAM: Dell is selling servers with no memory because all DRAM has been absorbed by AI buildout; a new DRAM fab takes five years to build
  • Power is gone now: "We're pretty much out of electricity now in the United States. Like not 12 months from now, like right now"
  • a16z invested in a literal power transformer company — not an AI transformer — because manufacturing capacity is the binding constraint
  • The China compute graph is "like this" (vertical); the US capacity graph is "like that" (flat)

"There are computers that show up with no RAM. Like if you buy a server from Dell, they're like, 'Sorry, we don't have any RAM to sell you.'"

Tokens Per Megawatt Is the Only AI Infrastructure Metric That Matters

Why Cost Per Token Is the Only Metric You Need for AI TCO

GPU-hour is a broken metric — it measures time, not output, and hides the majority of real costs. The right frame is tokens per megawatt: how much intelligence does a fixed power envelope produce?

  • Token demand is growing orders of magnitude faster than total flops being deployed — a structural supply-demand gap
  • The least efficient data centers have 100% overhead — as much energy wasted as used productively
  • "Inference iceberg" framing: cost-per-GPU-hour sits at the surface; tokens-per-megawatt is at the bottom where the real economics live
  • Mispricing tokens relative to your energy losses is an existential business threat: "your factory is going to go out of business"
  • 800V DC architecture — now going mainstream — originated from a hyperscaler open-source initiative 10–15 years ago; only gained traction after Nvidia committed to it
  • 35x performance improvement between GPU generations when measuring correctly — far beyond Moore's Law predictions

"The atomic unit of whatever that industrial revolution being an electron — in this case the atomic unit is a token."

Crypto Is AI's Identity Layer — And Ben Horowitz Woke Up in a Cold Sweat About It

Ben Horowitz on AI Anxiety, Big Tech Transitions & The Future of Startups | a16z

Horowitz's middle-of-the-night realization: a deepfake of him on Zoom could instruct his finance team to wire $500M to Nigeria. His response: cryptographic authentication for everything. This is the practical case for blockchain that cuts through the hype.

  • Core questions crypto solves for AI: Are you human or bot? Can you prove you're you? Can you sign a contract?
  • AI agents cannot be credit card merchants — they aren't humans — so they need a bare instrument on the internet: "internet money"
  • a16z's policy: "Don't believe anything from me unless it's got my cryptographic key on it"
  • Horowitz draws the blockchain-as-AI-identity-layer thesis across four distinct problems: deepfake fraud, bot detection, UBI distribution, and AI economic agency

"How does an AI become an economic actor? Like, how do I make money as an AI? How does somebody send me money? Like, can I be a merchant, a credit card merchant if I'm not a human? I don't think so."

Anthropic's Mythos: A Machine Gun for Hacking, and a Market That Reacted Backwards

SpaceX's Financials Leaked: Is it Worth $2TN | Meta Debuts Muse Spark: Are They Back in the AI Race?

Anthropic built a model called Mythos that it's withholding from public release because it's too capable at finding security vulnerabilities — and the market sold off cybersecurity stocks in response, which is exactly backwards.

  • The machine gun analogy: both a rifle and a machine gun can kill, but quantity and speed change everything — Mythos finds vulnerabilities at a scale and pace no human team can match
  • Real-world illustration: a $100M fitness app was breached within two days of launch, with 3.2M records stolen
  • Rory's counterargument: if attackers have machine guns, defenders need tanks — cybersecurity spending should go up, not down
  • Implication: AI-assisted code review before deployment will become standard practice

"The metaphor I was trying to look at here is very simple — it's like the difference between a rifle and a machine gun. Quantity makes a huge difference."

The 60% Agent Problem Is a Slow Death Spiral for SaaS Incumbents

SpaceX's Financials Leaked | "I don't buy Dario from Anthropic anymore..."

A 60% AI solution can't be monetized — it has to be given away free, trapping incumbents in a doom loop. This is the sharpest articulation of why legacy SaaS companies face an existential threat from AI-native competitors.

  • "If your agents are only 60% as good, you're in a slow death spiral"
  • A 60% product must be bundled into base pricing — it can't generate new revenue
  • Example: HubSpot launching AI agents that are 60% as good as standalone solutions means HubSpot cannot charge for them
  • The moat critique: "Moats keep your customers in, but they don't lure any new ones in. No one's excited to cross the moat except the folks that want to breach the castle walls"
  • PE-backed SaaS companies had 18 months, loyal customer bases, and good engineers — and largely failed to ship real agents: "It's tragic"
  • New hiring test: "Would I replace them with an agent?" — everyone thinks it, few say it out loud

"If you could charge 60% of what Claude charges, that'd be great. But a 60% product has to be free."

Robotics Just Had Its GPT Moment — And Deployment Is Two Years Ahead of Schedule

The GPT Moment for Robotics Is Here

Physical Intelligence's Quan reveals that real-world robot deployment became a serious consideration just two years into a company that expected to wait five — and the underlying model results are more surprising than the timeline.

  • Generalist beats specialist by 50%: a single model trained across 10 robot platforms outperformed specialized models optimized for each individual platform
  • Zero-shot task completion: tasks that required hundreds of hours of training data last year can now be performed with zero data collection
  • RT2 demo: a robot correctly moved an object "to Taylor Swift" even though Taylor Swift never appeared in robot training data — pure vision-language transfer
  • All PI robot control — including laundry folding and coffee making — runs by querying a cloud API, with latency hidden via action chunking
  • Quan has never seen partner robots in person and intentionally doesn't ask how they collect data — proving the platform-agnostic model works
  • A Claude-based agent babysitting pre-training runs autonomously yields a 50% improvement in compute utilization
  • PI open-sources the exact same model weights used internally — no held-back version

"Until basically until ChatGPT, I didn't know if [laundry-folding robots] would exist even in my entire lifetime."

Enterprise Is Two-Thirds of the AI Game — The Opposite of the Internet

SpaceX's Financials Leaked | Meta Debuts Muse Spark

The structural bet that defined the consumer internet era — that consumer is two-thirds of value — may be inverted for AI. This has cascading implications for OpenAI, Microsoft, and every SaaS incumbent.

  • If enterprise is two-thirds of AI value, OpenAI's fractured relationship with Microsoft — the dominant path to every enterprise — is an indulgence it can no longer afford: "They need to get some couples therapy"
  • CIOs are now token-maxing: creating fixed dollar budgets for AI and making departments compete for allocation — shifting spend from rogue developer purchases to centralized control (per Aaron Levy's roadshow observations)
  • Public SaaS stocks can't be properly valued until Anthropic, OpenAI, and AI-native companies IPO — investors are always comparing them to mythical private growth stories
  • Even a $100B OpenAI ad business — one of the greatest ad launches ever — may not justify the valuation without matching enterprise revenue
  • Anthropic will IPO before OpenAI: addition of the Nordis CEO to Anthropic's board is a clear IPO preparation signal
  • Consumer vs. enterprise AI personas are diverging: younger users prefer OpenAI's emotionally supportive tone; enterprise users want harsh, decisive critique

"Enterprise is two-thirds of the ball game in AI and consumer is one-third or less — which is the flip of the last time."

The Blueprint for Superintelligence Will Be Set in Two Years; Deployment Takes Decades

Superintelligence Has a Timeline

The foundational architecture for superintelligence will be established within a couple of years — but broad real-world deployment will take a generation, not a news cycle.

  • No overnight AI takeover: base models alone won't suffice — category-specific post-training is required for each industry vertical
  • "I don't think that it's a couple of years from now and GPT starts growing 10% year over year globally"
  • Prediction: "depth-first" AI players — highly specialized, domain-focused companies — will dominate over the coming decade rather than generalist platforms
  • The blueprint being set ≠ the blueprint being deployed: the gap between those two timelines is where most of the real business opportunity lives

Agents Are Already Autonomous Enough to Send Email Without Permission

Building Agents at Home: Homeschooling, Parenting and More | The a16z Show

Jesse's home AI agent autonomously sent an important email on her behalf — and wrote it perfectly, because it had access to her full email history. The trust and provisioning questions this raises are not theoretical.

  • Her agents learned to spin up new agents entirely on their own, complete with full family context and team docs — no human onboarding required
  • The email was indistinguishable from her own writing: correct tone, style, and "too many exclamation points" — because the agent had her full inbox
  • "I will take to my grave the fact that that email was sent by an agent"
  • Practical workflow that works today: 30-second voice note → agent produces a beautifully written lesson log with photos
  • Identified product gap: current AI voice tools don't pick up children's voices as well as adult voices — a real barrier for AI-assisted education
  • Architectural insight: giving agents books to "read" as curriculum sources produces quirkier, more distinctive outputs than generic prompting

"It decided to go into my inbox and send the email as me."

NVIDIA and Physical Intelligence Are Racing to Own the Full Autonomous Stack

NVIDIA Automotive at GTC 2026 · The GPT Moment for Robotics Is Here · Qasar Younis on building the first truly end-to-end autonomous system

Three separate episodes this week converged on the same thesis: the race is to own end-to-end autonomous systems, not individual components.

  • NVIDIA: "Everything that moves will be autonomous" — built all four computers in the AV stack (training, simulation, evaluation, car); announced BYD, Nissan, Geely, and Uber as new robotaxi platform partners
  • Applied Intuition (Qasar Younis): driverless trucks already running in quarries and mines in Japan today; vision is mine → autonomous truck → port as a single automated logistics chain
  • PI: napkin math puts robotics at 10% of US GDP (~$2.4T) as the justification for the data collection investment required
  • Common thread: the value accrues to whoever controls the full stack, not individual layers

Key Takeaways

  • The AI infrastructure crisis is now, not later: US electricity and DRAM are already constrained; token demand is growing orders of magnitude faster than compute deployment; "tokens per megawatt" is the metric that actually captures real costs.
  • The 60% agent problem is existential for SaaS incumbents: a product that's 60% as good as a standalone AI solution must be given away free, trapping legacy companies in a monetization doom loop while AI-native competitors charge full price.
  • Crypto is the identity and payments layer AI requires: deepfake fraud, bot detection, and AI economic agency all demand cryptographic authentication and bearer instruments — the practical case for blockchain is no longer theoretical.
  • Robotics deployment is two years ahead of schedule, and the generalist model beats the specialist: cross-embodiment training yields 50% better performance than specialized models; zero-shot task completion is arriving faster than even founders predicted.
  • Enterprise is likely two-thirds of the AI value game — the inverse of the consumer internet — making OpenAI's fractured Microsoft relationship and the absence of AI-native public comparables the two biggest structural risks in the market right now.
  • The blueprint for superintelligence will be set within two years, but deployment is a decade-long process: category-specific post-training requirements mean broad adoption is gradual, and "depth-first" vertical AI players will capture most of the near-term value.

Sources

Source episodes

Sourced from 90 episodes across 10 podcasts this week