An Unlikely Consensus
When Anthropic CEO Dario Amodei published his manifesto, “We Must Pace the Frontier,” on September 12, 2026, the tech world witnessed something rare: immediate, public alignment among bitter rivals. Amodei argued that recursive self-improvement and agent swarms demand an intentional slowdown. Within hours, OpenAI’s Sam Altman backed the premise, and Elon Musk added his endorsement.
When Anthropic CEO Dario Amodei published his manifesto, “We Must Pace the Frontier,” on September 12, 2026, the tech world witnessed something rare: immediate, public alignment among bitter rivals. Amodei argued that recursive self-improvement and agent swarms demand an intentional slowdown. Within hours, OpenAI’s Sam Altman backed the premise, and Elon Musk added his endorsement.
Amodei’s technical concerns around safety, rogue agent swarms, and internal sandboxes may be entirely legitimate. But I don’t have a window into Anthropic’s, OpenAI’s, or xAI’s private clusters—and outside a tiny circle of researchers, neither does anyone else.
While we should take frontier safety seriously, another explanation is worth considering: What if the call to “pace the frontier” isn’t just about catastrophic risk? What if it’s a convenient narrative cover for the fact that foundational progress is hitting diminishing returns, enterprise adoption is stalling, and commercial monetization is moving far slower than projected right before massive liquidity events and IPOs?
If that alternative holds any weight, it validates a reality that field-tested operators have understood for years: the industry’s obsession with horizontal AI was an expensive detour.
Exhaustion of the Spaghetti Cannon
For the past three years, the tech sector has operated a multi-hundred-billion-dollar spaghetti cannon, firing generalized foundation models horizontally across every imaginable sector to see what sticks.
The industry wrapped foundational models in lightweight SaaS interfaces, slapped “copilot” on enterprise pricing tiers, and promised immediate productivity revolutions to boardrooms everywhere.
The results are now visible across enterprise balance sheets:
- Enterprise generative AI initiatives have hit a wall of operational exhaustion. More than half of corporate AI pilots stall after proof-of-concept, abandoned due to unpredictable inference costs, unmitigated hallucinations, and an inability to demonstrate tangible financial return.
- The capital expenditures committed by hyperscalers to secure gigawatts of power, data centers, and advanced silicon continue to outpace the actual top-line revenue generated by enterprise software applications by hundreds of billions of dollars.
- Beyond digital marketing copy and meeting summarization, actual integration of autonomous models into revenue-generating, physical-world operational workflows remains stuck in single-digit percentages across global business.
Horizontal AI treats every industry as if it were a digital marketing department. But heavy enterprise doesn't run on polite text generation. It runs on physical tolerances, legacy hardware, strict regulatory liability, and zero-error-margin operating engines. Firing generic algorithms against those environments didn't build an industrial revolution; it created market friction and burned through billions in enterprise goodwill.
Cloud Ideology vs. Physical Reality
The debate over the pace of AI exposes an ongoing philosophical divide across tech leadership:
On one side are the Cloud Theorists, led by figures like Amodei and Altman. Their worldview is built on scaling laws, centralized data centers, and the conviction that emergent software intelligence in the cloud will cascade downward to fix real-world problems. When they urge the world to slow down, they frame it around managing an emergent digital intellect.
Yet when valuations depend on maintaining astronomical multiples ahead of anticipated public offerings, admitting that the current architecture has hit practical hardware limits, data exhaustion, or enterprise pricing pushback is financially untenable. Framing deceleration around responsible stewardship and global safety is clean, defensible, and universally praised by regulators.
On the other side are the Physical Pragmatists, reflected in the industrial doctrines of leaders like Elon Musk and Palmer Luckey. Their thesis starts from an unforgiving premise: language is easy, but physics is hard.
Musk’s focus on real-world perception networks and Luckey’s deployments of edge-first defense hardware at Anduril demonstrate that software cannot exist in a vacuum. A model living on an AWS server cannot navigate kinetic environments, survive electronic warfare, or manipulate heavy machinery without bespoke physical hardware, local silicon, and deep vertical engineering.
Whether the frontier labs are genuinely terrified of their creations or simply buying time to patch their economic balance sheets, the end result is identical: the horizontal cloud narrative has run out of road.
Twelve Years in the Dirt
The alternative to the horizontal spaghetti cannon is vertical integration: going deep into a specific industrial domain, understanding its physical realities, and building custom intelligence that solves an expensive, measurable failure point.
It is an agonizingly slow, capital-intensive path.
When we built Farmwave, our objective wasn't to generate harvest poetry or chat with an agronomy database. It was to build an edge-compute computer vision system mounted directly to the chassis of combine harvesters to mechanically measure and reduce the 2% to 5% of global yield routinely lost out the back of the machine.
That engineering reality looked nothing like a Silicon Valley software sprint:
- You cannot train models on public internet scrapes; you must spend twelve harvest seasons in the field across continents, capturing and manually labeling millions of ground-truth images of cracked kernels and crop conditions under variable sunlight and abrasive conditions.
- You cannot rely on cloud connectivity when operating across vast rural fields. Inference must happen locally, running sub-second calculations on ruggedized edge processors exposed to extreme vibration, dust, and 115-degree heat.
- You have to earn the trust of machine operators whose livelihoods depend on machine uptime during narrow weather windows.
That process required twelve years and fifteen million dollars.
Silicon Valley’s venture funding cycle—predicated on 18-month software iterations, rapid customer acquisition, and zero marginal costs—is structurally unsuited for that kind of development. But this is how enduring value is created. Deep vertical AI requires capital, domain immersion, and hardware durability. When implemented, the economic return isn't a vague percentage bump in "productivity"—it is millions of dollars in recovered commodity yield and an unassailable data moat.
The Real Discovery Point
The tech industry may have arrived at an overdue reckoning. The marketplace has reached a discovery point: enterprise customers are refusing to pay premium subscriptions for thin wrappers, compute budgets are colliding with energy infrastructure, and generic models cannot bridge the gap to industrial utility.
If leading labs are indeed tapping the brakes because capability gains are hitting friction rather than running away into superintelligence, it changes the strategic landscape for everyone else:
- Enterprises can stop redesigning their software stacks every quarter in terror of the next foundation model release. Pacing stabilizes the foundation, allowing leadership to focus on solving actual structural problems.
- Capital spent subsidizing centralized cloud compute can redirect toward edge silicon, sensor integration, robotic actuation, and domain-specific dataset acquisition.
- The era of prompt engineers presuming to optimize operations they have never physically observed is ending. Value returns to the teams who understand the nuance, friction, and mechanics of the vertical industry they serve.
Ground Truth Over Speculation
Dario Amodei suggests we pace the frontier so humanity can survive the fire we are kindling. Perhaps that threat is real. Or perhaps the industry is simply running out of runway to justify its current commercial promises, using safety as a respectable holding pattern while figuring out what customers actually want.
Dario Amodei suggests we pace the frontier so humanity can survive the fire we are kindling. Perhaps that threat is real. Or perhaps the industry is simply running out of runway to justify its current commercial promises, using safety as a respectable holding pattern while figuring out what customers actually want.
Either way, a deceleration is welcome.
If the frontier slows down, the market can finally stop throwing algorithms against the wall to see what sticks. Capital can stop chasing synthetic benchmarks and start doing the patient, gritty engineering that transforms raw technology into durable industrial infrastructure.
Real-world problems are not solved in a cloud sandbox. They are solved in the mud, on factory floors, in supply chains, and on the mechanical iron that keeps civilization running.
Let the frontier pace itself. It’s time for artificial intelligence to grow up and get to work.
Craig