AI Sovereignty: Can we trust Sam and Dario?

“I’m not throwing shade at them, but something has gone completely wrong. The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens. I’m going to get no value and they’re going to get my IP.”

— Alex Karp, Palantir CEO, CNBC Squawk Box, July 1, 2026

“These people are stealing the weights and alpha of my business, and they’re creating a wealth tax.”

— Alex Karp, same interview

He wasn’t talking about abstract AI safety. He was talking about Sam and Dario — by first name. The CEOs of OpenAI and Anthropic. Leaders I have followed closely for years, respected for what they’ve built, and watched evolve in real time as the stakes changed around them.

Three weeks earlier, a room full of optimists at Sequoia Capital’s AI Ascent conference had made a different declaration entirely: functional AGI has arrived. The cars are here. Services is the new software. Whatever you imagined building over the next hundred years is now possible in a hundred days.

Both things can be true at once.

The wave is real. I’ve seen it on the ground — once you see it, it’s hard to unsee it. If you’ve read my Reflections on AI series, or my 2025 year-in-review, you know I’ve been tracking this arc for a while. But we need to walk into this moment eyes wide open. Things are unfolding in real time for everyone. There is no chapter in the playbook for what happens when the most capable technology in human history arrives inside a business model that may structurally conflict with the people paying for it.

This post is my attempt to hold both truths: the exhilaration of the computation revolution, and the erosion of trust in who gets to own it.

Act I: This Is AGI — Sequoia’s Frame

A few weeks before Karp’s CNBC appearance, I watched Sequoia Capital’s AI Ascent 2026 keynote — “This is AGI.” Pat Grady, Sonya Huang, and Konstantine Buhler didn’t deliver a product launch. They delivered a reframe. If you want to understand why Karp is so angry — and why Apple sued OpenAI the following week — you have to understand what Sequoia declared first.

Three Reasons This Wave Is Different

Sequoia argued this AI cycle is unlike internet, cloud, or mobile in three specific ways.

1. The biggest wave yet. Not just software — software and services. They put a number on it: roughly $10 trillion in addressable services labor. Legal services in the United States alone is on the order of $400 billion — about the size of the entire global software industry. AI isn’t replacing your CRM. It’s replacing the work your CRM was built to support.

2. The fastest wave yet. Adoption curves that took a decade are compressing into months. The floor moves underfoot daily. New model capabilities drop while you’re still integrating last quarter’s API.

3. A different kind of wave — and this is the insight that stuck with me. There are two basic kinds of technology revolutions. Communication revolutions change how information is distributed — the printing press, the telephone, the internet, mobile, social media. Computation revolutions change how information is processed. AI is the latter. We haven’t lived through a computation revolution at this scale in our lifetimes.

That’s why the metaphor isn’t “faster horses.” It’s cars. And Sequoia’s position is blunt: the cars have arrived.

A Computation Revolution, Not a Communication One

I’ve lived inside technology waves my entire career — client/server, the internet, cloud, mobile, social. Each one was real. Each one reshaped how we work. But I have never seen anything like this in my lifetime.

For decades, my mental model of progress was governed by Moore’s Law and its intersection with traditional CPUs. More transistors, more clock cycles, more cores — predictable, incremental, measurable. You could plan around it. When I was a kid building desktop machines, the cases often shipped with a gimmicky LED readout on the front panel showing the CPU clock speed in MHz — a fun, slightly absurd reminder that next year’s chip would make this year’s number obsolete. The waves I listed above were mostly communication revolutions: faster distribution of information, better interfaces, cheaper reach. Important. Transformative. But the substrate underneath stayed familiar.

The closest I came to a computation shift before this was at DigitalGlobe around 2010 — a story I told briefly in Artificial Intelligence & Me — where I helped build one of the first, and at the time one of the largest, GPU-based supercomputers for processing digital satellite imagery. That’s where I learned CUDA — NVIDIA’s parallel computing platform, still called CUDA today, though the stack around it has grown enormously (cuDNN, TensorRT, NIM). GPUs weren’t the default compute path; we were repurposing hardware built for graphics to run algorithms that classical CPUs couldn’t touch at scale — cloud cover detection, change detection, image feature recognition. It was a genuine leap. But the applications were narrow: satellite imagery, geospatial intelligence, defined inputs in a defined domain. Powerful, yes. General, no.

What Sequoia is describing now is a computation revolution without that boundary. The substrate itself is shifting — from CPUs executing instructions to models reasoning over context, from software you configure to agents that persist and iterate. I felt something similar watching GPU clusters chew through terabytes of orbital imagery. This feels like that moment, except the domain is everything.

Long-horizon agents are Sequoia’s functional definition of AGI: dispatch an agent, it recovers from failure, persists until the job is done. ChatGPT (2022) gave us knowledge. o1 (2024) gave us reasoning. Claude Code and peers (2025–2026) gave us iteration over time.

METR (Model Evaluation and Threat Research) has been tracking what they call the 50% time horizon: the length of real-world tasks — measured in how long they take a human professional — that a frontier model agent can complete autonomously with 50% reliability. Plot it on a log scale and the curve is stubbornly exponential. From 2019 through 2025, the horizon doubled roughly every seven months. A task that took an agent minutes in 2022 stretches toward hours today; extrapolate the trend and you’re looking at day-long expert work by 2028 and year-long projects not long after. METR’s updated Time Horizon 1.1 analysis suggests the post-2024 pace may be even faster — closer to three months per doubling. Even if the absolute numbers are wrong by an order of magnitude, the direction is the story.

If that’s right, the implications are staggering. Not 10–40% productivity gains. 10–40x. Not copilots. Colleagues. Not software you license. Work you buy.

Quantum Computing: Another Kind of Computational Wave

Around the same time I was digesting this keynote, I was taking MIT’s Introduction to Quantum Computing — the same season I was writing about latent space — and the parallel clicked in a way I didn’t expect. Quantum computing isn’t AI. Different physics, different engineering constraints, different time horizon for mainstream adoption. But it rhymes with what Sequoia is describing: a shift in the substrate of computation itself, not just a faster way to send messages.

Classical computing processes information as bits — zero or one. Quantum computing exploits superposition and entanglement to explore vast solution spaces that classical machines cannot efficiently traverse. Problems in cryptography, molecular simulation, and optimization that are intractable at human timescales become addressable. You aren’t building a better router. You’re building a different kind of processor for a different class of problem.

When the substrate shifts, friction shows up in places incumbents never had to defend: data ownership, model weights, the ephemeral layer of insights generated on top of your core assets. New physics, new politics.

Services Is the New Software

The keynote’s economic punch line: services is the new software.

Konstantine Buhler’s cognitive revolution framing is the one that keeps me up at night. The Industrial Revolution mechanized physical work — today, something like 99% of physical labor is done by machines, not human muscle. The AI wave targets cognitive work — and Sequoia’s bet is that 99.9% of it will follow, on a dramatically compressed timeline.

The aluminium analogy lands hard: aluminium was once more precious than gold. The tip of the Washington Monument is aluminium — a statement of wealth. Then electrolysis made it cheap. PhD-level cognitive skills, Sequoia warns, may follow the same curve. Skills we treat as precious today become disposable tomorrow.

Some of the most valuable human capabilities in the economy are on a commoditization path. That’s not dystopian fiction. That’s the investment thesis behind the most important technology conference of the year.

And here’s where the friction starts. If services is the new software, who owns the service layer? For two decades, enterprises bought software and kept their processes, data, and domain expertise in-house. The software was the tool. Now the tool can do the work. The frontier labs aren’t selling picks and shovels. They’re selling the miners — and learning from every seam they dig.

Act II: The Fracture — This Past Week

Sequoia painted the opportunity. Then July happened.

Three stories landed in the same news cycle — each a different face of the same question: who owns the means of production in the age of functional AGI?

Karp Goes Off-Script

On July 1, Alex Karp sat down with CNBC’s Squawk Box to announce Palantir’s expanded partnership with NVIDIA — a Sovereign AI Operating System built around open-weight models (including NVIDIA’s Nemotron family) for government and enterprise deployments where data, hardware, and model weights stay under customer control.

What followed was less a CEO interview and more a twenty-minute monologue. Karp named Sam Altman and Dario Amodei directly. He said enterprises are “livid.” He called token-based pricing a “wealth tax.” He asked whether America should “outsource the battlefield of this country to the consensus view in Silicon Valley” and answered his own question: “That is effing insane.”

When Becky Quick told him he sounded angry, Karp said: “This is the voice of American business that is being channeled through me.”

Palantir had already published a nine-point AI sovereignty manifesto on X — framing data retention as treasure, token-maxxing as a broken business model, and ownership of compute, models, and data stack as non-negotiable. Karp’s CNBC appearance was the manifesto with the volume turned up.

I don’t agree with everything Palantir sells. Karp has obvious commercial incentives. But dismissing him as a competitor throwing shade misses the point. He’s channeling a sentiment I hear in rooms that don’t make CNBC — from operators writing eight-figure AI checks who can’t tell you what the ROI is.

Sacks, Figma, and the Duopoly Frame

Two days later, David Sacks made the same argument in a different register.

On X and then on the All-In Podcast, Sacks said Karp was “exactly right.” His framing: for enterprises, “AI safety” isn’t abstract alignment research. It’s control over compute, models, data stack, and alpha — the proprietary knowledge that makes your business yours.

Sacks pointed to Anthropic’s launch of Claude Design as the cautionary tale. An Anthropic executive had served on Figma’s board. Then Anthropic shipped a design tool that competed directly with a key partner’s core product. Figma’s stock had fallen roughly 50% year-to-date while Anthropic’s valuation surged. Whether you call that fair competition or a trust violation depends on where you sit. If you’re an enterprise technology leader watching your AI vendor verticalize into your business, it feels like the latter.

Sacks went further: OpenAI and Anthropic have formed an effective duopoly at the frontier model layer. Anthropic’s push for safety regulation, in his view, risks regulatory capture — rules that cement incumbents rather than protect the public.

Dario would push back — and fairly. Anthropic’s regulatory proposals target models above 10²⁵ FLOPs and labs above $500M in AI revenue. This isn’t a dragnet on garage startups. Genuinely dangerous capabilities in cyber and biosecurity deserve serious conversation. The enterprise question is different: whether safety advocacy and vertical product strategy can coexist with partner trust when the same company sells you the API and ships your competitor.

The Information’s reporting on Anthropic blindsiding business partners gave the narrative institutional weight. Partners who bet their roadmap on API access woke up to a competitor.

A few days later, Sacks pressed the same thread on All-In — and named the problem enterprises rarely say out loud in a vendor meeting:

CTOs want to shift token spend to cheaper open-weight models. Compute bills are skyrocketing, and the sovereignty anxiety Karp articulated is real — nobody wants to hand proprietary alpha to a lab that might ship their competitor next quarter. But most enterprises can’t execute the shift. Coinbase and DoorDash are the exceptions that prove the rule: they built token-routing layers that send frontier tasks to frontier models and everything else to cheaper open or distilled tiers. Your average Fortune 500 doesn’t have that plumbing yet.

The wallet-share data is sobering. Closed-model spend isn’t losing ground — open-weight models reportedly fell from roughly 19% to 11% of enterprise AI wallet in a year, even as total usage explodes. Part of that is a measurement artifact (self-hosted inference shows up as compute, not token bills). Part of it is capability. So the duopoly isn’t only regulatory capture or partner betrayal. It’s default gravity: the path of least resistance still leads to Sam and Dario, even when everyone in the room agrees with Karp.

Apple Sues OpenAI

If you needed proof that trust fractures aren’t limited to angry CEOs and podcast hosts, Apple provided it on July 10. The New York Times reported that Apple filed a federal lawsuit accusing OpenAI of trade secret theft — alleging that OpenAI, two former Apple employees (including hardware chief Tang Tan and engineer Chang Liu), and Jony Ive’s io Products engaged in a coordinated effort to extract confidential product designs, manufacturing processes, and supply chain strategies. Apple claims more than 400 former employees now work at OpenAI. It alleges interviewees were coached to bring “actual parts” from Apple to OpenAI interviews.

This is the company that partnered with OpenAI in 2024 to integrate ChatGPT into Siri — now suing its partner for misappropriating the very hardware expertise Apple spent decades building.

The partnership isn’t formally at issue in the suit. But the symbolism is brutal. Reports that OpenAI considered suing Apple over Siri integration — alleging Apple hadn’t promoted ChatGPT as promised — suggest the relationship was rotting from both ends before the trade-secrets complaint landed. If Apple — with its legal budget, its leverage, its negotiated agreements — can’t trust the relationship, what should a mid-market enterprise with a standard API agreement expect?

Three stories. Same week. Same question.

Act III: The Trust Audit

Trust erodes when expectations and incentives misalign. Here’s what enterprises are actually buying — and what they may be giving up.

What Is a Frontier Model?

A frontier model is a large-scale foundation model at or near the state of the art — trained on massive compute budgets, updated frequently, capable of general reasoning, coding, tool use, and long-context work. Today, the hosted frontier layer is dominated by a small set of labs: OpenAI (GPT-4.x / o-series / GPT-5.x), Anthropic (Claude Opus/Sonnet), Google (Gemini), Meta (Llama — open weights, though Meta-hosted services exist), and a handful of challengers.

OpenAI and Anthropic are the focus here because they are the default enterprise choices — and because they are building vertical applications on top of the API layer they rent to you. ChatGPT, Claude Code, Claude Design, operator agents, coding agents, design tools. The API is the beachhead. The application is the moat.

Should You Assume Frontier Models Will Come for Your Business?

Yes — not out of malice, but out of incentive structure.

If services is the new software, the frontier labs are not content to be infrastructure forever. Every API customer is a market research study. Every fine-tuning job reveals domain structure. Every enterprise deployment teaches the lab what “good” looks like in legal, finance, design, coding, support, security.

The pattern is consistent:

1. Release a capable general model.
2. Rent access via API at token prices.
3. Learn from usage what verticals have the highest willingness to pay.
4. Ship a first-party application that captures the value chain.

You are not just a customer. You are a training distribution channel for their product roadmap.

Is Your Data Really Safe?

“Your data is not used for training” is the enterprise sales line. Read the fine print. Consider:

• Retention policies — how long are prompts, completions, and logs stored?
• Subprocessors — who touches your data in the inference chain?
• Enterprise vs. consumer tiers — different terms, different enforcement
• Derived artifacts — embeddings, eval sets, fine-tuned adapter weights — who owns them?
• Legal jurisdiction — where does inference run, and whose law applies?

Even with strong contractual terms, you are sending proprietary context into a system you do not control, running on hardware you do not own, governed by a usage policy that can change with a blog post.

Sam and Dario are thoughtful leaders. I believe they take safety seriously. But safety and sovereignty are not the same thing. You can build a safe system that still commoditizes your customers.

Token Economics and the ROI Gap

Karp’s “wealth tax” is crude but directionally fair. Token pricing charges you for activity, not outcomes. A failed agent run costs money. A hallucinated draft costs money. A support bot that escalates everything costs money.

Most enterprises I talk to cannot answer a basic question: what is our cost per successful task? They know monthly API spend. They don’t know cost per resolved ticket, per generated contract, per analyzed dataset. Without that, ROI is a faith exercise — and faith is eroding.

Guardrails: Great for Commodity, Oppressive for Deep Science

Frontier model guardrails exist for good reason. Refusal behaviors on bioweapons, exploitation, and fraud protect the labs and their users.

But guardrails are trained on consensus distributions. Deep science, novel engineering, specialized legal reasoning, and frontier research often look like edge cases to a safety layer optimized for consumer chat. I’ve watched researchers fight the model to do legitimate work — not because the model can’t, but because the harness won’t let it persist long enough or explore far enough. There’s a parallel here to what I wrote about in The Stochastic Era: the model’s probabilistic nature is a feature until your work lives in the tail of the distribution.

If your competitive advantage lives in the tail of the distribution, hosted frontier guardrails are not neutral. They are a ceiling.

The Ephemeral Layer — Who Owns the Insights?

Here is the question I think gets too little airtime.

When you run AI against your core data — customer records, operational telemetry, support transcripts, field notes, financial models — you generate a second layer of value. Summaries. Embeddings. Cluster labels. Anomaly flags. Recommended actions. Agent trajectories. Fine-tuned LoRA weights.

Call it the ephemeral layer — intelligence derived from your gold, sitting on top of your gold.

Consider an anonymized example. A national retailer deploys AI across its call centers — thousands of conversations per day about product defects, shipping failures, pricing confusion, competitor mentions. The core data is the call recordings and CRM records. The ephemeral layer is everything the model learns:

• Which product SKUs generate the most unresolved frustration
• Which competitor names surface before churn
• Which agent scripts work and which don’t
• Emerging defect patterns weeks before they hit the warranty database

Who owns that layer?

If inference runs on a hosted frontier API, the answer is murky. Your contract may cover raw prompts. It almost certainly doesn’t cover what the model inferred from a million calls — inferences the vendor’s systems processed, logged, and potentially retained. The insights may be more valuable than the transcripts.

Your data is gold. The ephemeral layer is the refined product. And right now, most enterprises are shipping both to someone else’s refinery.

Act IV: AI Sovereignty — The Alternative Path

“Sovereignty” sounds geopolitical. In practice, it means control over the parts of the stack that generate alpha:

Layer	Sovereign question
Data	Does my proprietary data leave my boundary?
Weights	Do I own or control the model parameters?
Inference	Who runs the GPUs, and under what terms?
Orchestration	Who owns the agent harness and tool integrations?
Ephemeral layer	Who owns derived insights, embeddings, and fine-tunes?

You don’t need to build GPT-5 in your basement. Sovereignty is a spectrum.

The Open-Source Model Landscape

First, kill a myth: open-source AI is not all Chinese.

The open-weight landscape in mid-2026 is genuinely global — and license terms matter more than geography.

United States

• Meta Llama 4 (Scout, Maverick, Behemoth) — the deployment default for ecosystem breadth. Llama 4 Scout pushes context windows to 10 million tokens in a MoE architecture (109B total / 17B active). Llama Community License — permissive with revenue thresholds to watch.
• NVIDIA Nemotron — optimized for enterprise inference via NIM microservices; the Palantir sovereign stack partnership runs here. Strong fit if you’re already on NVIDIA hardware.
• Google Gemma 4 — efficient open models (e.g., 26B A4B variants) for local and edge deployment under Gemma Terms of Use.
• Microsoft Phi-4 — 14B-parameter model with surprising reasoning for its size; MIT license; good single-GPU target.
• OpenAI GPT-oss 120B — OpenAI’s open-weight release (Apache 2.0); practical single-H100 deployment for teams that want OpenAI-ish capability without the API dependency.

Europe

• Mistral (Large 3, Mixtral, Codestral) — France-based; Apache 2.0 on most releases; strong multilingual performance; MoE efficiency (e.g., Mixtral 8x7B activates ~13B of 47B parameters per token). The cleanest permissive-license path for regulated European enterprises.

Canada

• Cohere Command R+ — retrieval-focused; strong RAG story; license is more restrictive (CC-BY-NC) — read before commercial deploy.

China (significant capability, varied licenses)

• DeepSeek (V3, R1, V4 families) — frontier-class reasoning and coding at aggressive price points; MIT or model-specific licenses on weight releases; self-host with SGLang/vLLM but ops maturity required.
• Qwen 3 / 3.5 (Alibaba) — Apache 2.0 on key releases; strong reasoning MoE models (e.g., 397B total / 17B active).
• Kimi K2.x (Moonshot) — MIT-licensed; agentic coding strength.
• GLM-5 / GLM-5.1 (Zhipu) — open-weight models competitive on long-horizon software engineering tasks.

How to choose (rules of thumb, not gospel):

Priority	Start here
Broadest tooling ecosystem	Llama 4
Cleanest commercial license	Mistral Large 3, Qwen 3.5, Phi-4, GPT-oss
NVIDIA enterprise integration	Nemotron + NIM
Reasoning / coding on budget	DeepSeek R1/V4, Kimi K2.x, GLM-5
Long-context document analysis	Llama 4 Scout, Qwen 3.5
Single-GPU / local pilot	Phi-4, Gemma 4 26B, GPT-oss 120B, distilled DeepSeek variants

If I were standing up a sovereign stack tomorrow, I’d start with Llama 4 or Mistral Large 3 for breadth and licensing clarity, vLLM for inference, LoRA fine-tuning on a few thousand high-quality domain examples, and obsessive task-level observability — before anyone on the team starts fantasizing about full pre-training.

The Sovereignty Stack — Technical Detail

For a technical reader new to this space, here’s how the pieces fit.

1. Inference engine — turns weights into an API.

• vLLM — default for production throughput; PagedAttention; wide model support
• SGLang — strong for MoE models (DeepSeek), structured generation, high concurrency
• TensorRT-LLM — NVIDIA-optimized; pairs with NIM for enterprise SLAs
• Ollama — local dev and prototyping; not your production fleet

2. Model format and quantization

• Weights ship in FP16/BF16 (large), GPTQ/AWQ (4-bit), or FP8/NVFP4 (NVIDIA-optimized)
• Quantization trades precision for memory — a 70B model that won’t fit on two GPUs at FP16 may run on one at 4-bit
• Rule: benchmark your actual prompts after quantizing; don’t trust generic leaderboard scores

3. Fine-tuning (where most “build your own” journeys should start)

You probably don’t need pre-training. You need adaptation:

• LoRA / QLoRA — train small adapter matrices on top of frozen base weights; fraction of full fine-tune cost
• Tools: Unsloth, Axolotl, Hugging Face TRL
• Data: 1,000–10,000 high-quality domain examples can shift behavior meaningfully
• Output: adapter weights you own — merge or hot-swap per tenant

4. RAG (retrieval-augmented generation)

• Embed documents into a vector store (pgvector, Pinecone, Weaviate, Qdrant)
• Retrieve relevant chunks at query time; inject into context
• Sovereign because your knowledge stays in your database — the model only sees what you retrieve

5. Agent harness

• The difference between a chatbot and a long-horizon agent is the harness — memory, tool routing, retry logic, compaction, eval loops
• Sequoia’s MAD framework: Modes (wrap the customer), Affordance (make it simple), Diffusion (close the gap between capability and adoption)
• This is where Palantir’s “ontology” argument lives — and where you can differentiate without owning a foundation model. It’s the same gap I called out in The Great Corporate Efficiency Reckoning: the harness — evals, permissions, memory, rollback — is where pilots become production.

6. Observability

• Track tokens per task, latency, success rate, escalation rate, human-edit distance
• Without this layer, you’re Karp’s “chillax and waste my time with tokens” customer

7. Token routing

• Match task difficulty to model tier — frontier models for hard problems, open or distilled models for the long tail
• Coinbase and DoorDash built this internally; most enterprises haven’t. It’s the difference between sovereignty as a podcast sentiment and sovereignty in production
• Start simple: classify tasks, route by confidence, measure cost per successful outcome, expand coverage as evals prove the cheaper path

Building Your Own Model — It’s Not What You Think

“Build your own model” sounds like training GPT from scratch. For 99% of enterprises, that’s the wrong ambition.

Level 0 — Prompt engineering + RAG
Cost: engineering time. Sovereignty: high for data, low for model. Good for: knowledge bases, support, internal search.

Level 1 — Fine-tune open weights (LoRA)
Cost: thousands to tens of thousands of dollars in GPU time. Timeline: weeks. Sovereignty: you own adapters + data pipeline. Good for: tone, format, domain vocabulary, specialized classification.

Level 2 — Distillation
Train a smaller model to mimic a larger teacher. Cost: moderate GPU + data curation. Good for: edge deploy, cost reduction, latency-sensitive paths.

Level 3 — Continued pre-training
Extend a base model on domain corpus (code, legal, medical, logs). Cost: significant — multi-GPU clusters for days/weeks. Good for: domains where base models are genuinely weak.

Level 4 — Full pre-training from scratch
Cost: tens to hundreds of millions of dollars. Data moat required. Good for: foundation model companies — not for your insurance company’s IT department.

Hardware reality check: a 7B–14B model runs comfortably on a single GPU; a 70B model typically needs multi-GPU or aggressive quantization; a 400B MoE routes through vLLM or SGLang on a small cluster. Rent before you buy — but own the architecture decision either way.

The builder’s mistake is skipping to Level 4 because it sounds impressive. The operator’s win is Level 1–2 with a sovereign inference stack — open weights, your data, your adapters, your observability, your ephemeral layer under terms you control.

Hosted Alternatives That Aren’t OpenAI/Anthropic

Sovereignty isn’t only self-host. It’s optionality:

• Together AI, Fireworks, Groq, Baseten — host open models with SLAs; you choose the weight file
• Databricks Mosaic AI — enterprise data plane + model serving in your VPC
• Cloud AI Foundry / Bedrock / Vertex — mixed proprietary and open catalogs; read data residency terms carefully

The goal isn’t zealotry. It’s negotiating leverage — and a fallback when the frontier lab ships your competitor.

Sam and Dario — Leaders in Real Time

I want to be fair to two people I have respected.

Sam Altman saw the scaling laws before most, bet the company on them, and dragged the industry forward. OpenAI’s releases set the pace everyone else responds to. He has also presided over a company that partners with Apple while allegedly harvesting its trade secrets, pursues an IPO while enterprise customers question token ROI, and builds consumer hardware that competes with the platforms that distribute its models.

Dario Amodei articulated safety risks with more seriousness than almost anyone in the industry. Anthropic’s Constitutional AI and interpretability work matter. He has also built a company that sells safety to regulators while shipping vertical products that blindside board-level partners.

I don’t think either is venal. I think they’re playing a game whose rules require commoditizing everyone else’s alpha to justify frontier-model economics. And I think they’re evolving in public — faster than enterprises can update their trust models.

The question isn’t whether they’re good people. It’s whether their incentive structure aligns with yours once services is the new software.

Eyes Wide Open

Sequoia is right: this is a computation revolution. The cars have arrived. Long-horizon agents are functionally AGI for most business purposes. The cognitive revolution will commoditize skills we treat as precious today.

Karp and Sacks are right, too: the business model layer is fracturing. Token taxes without ROI. IP flowing to model weights. Partners becoming competitors. Ephemeral insights vanishing into vendor logs.

Apple’s lawsuit is right in a third way: when the stakes are high enough, trust breaks into the open.

None of this means the results on the ground aren’t real. They are. I’ve seen agents do work that would have taken teams days. I’ve seen models reason through problems that would have stalled organizations for weeks. Once you see it, it’s hard to unsee it — the same feeling I tried to capture in Context and Memory: The Gateway to AGI when the capability first became undeniable.

But seeing clearly means holding both truths:

1. Ride the wave — the capability is transformative and the cost of ignoring it is existential.
2. Own what matters — data, weights, inference, harness, ephemeral layer — or accept that someone else will.

AI sovereignty isn’t nationalism. It’s architecture. And for enterprise leaders watching Sam and Dario evolve in real time, it’s the question underneath every API key you issue:

Can we trust them with our gold?

I’m still writing the answer. So is everyone else.

-rjm