---
title: "Part I: Genesis"
part_of: "The Commanding Heights"
part_of_url: https://commanding-heights.com/commanding-heights.md
authors: ["Will Manidis", "Joe Roberts"]
date: 2026-10-29
url: https://commanding-heights.com/#part-i
summary: "The Genesis of intelligence is composed of the industry that makes and serves models, and it will keep spending on itself even as returns fall. Rational capital allocation at the level of the lab or hyperscaler can destroy capital in the aggregate. Intelligence’s tremendous revenues still come from early adopters deploying models into existing workflows. As with the railroads, the larger fortunes will be captured outside Genesis, by those who control how models reach the far larger economy they create."
---

## Part I: Genesis

> *The Genesis of intelligence is composed of the industry that makes and serves models, and it will keep spending on itself even as returns fall. Rational capital allocation at the level of the lab or hyperscaler can destroy capital in the aggregate. Intelligence’s tremendous revenues still come from early adopters deploying models into existing workflows. As with the railroads, the larger fortunes will be captured outside Genesis, by those who control how models reach the far larger economy they create.*

They are arguing about the AI bubble on X, the everything app. The bears can’t stop comparing hyperscaler capital expenditure to the fiber overbuild. The bulls point to Jevons and this cycle’s utilization rates. The models are busy proving open math questions.[^16] The mathematicians themselves are becoming luddites; even Terry Tao is foom posting. Staffers across Washington are asking Microsoft Copilot what post-training is while the President is handing awards to Elon, Brin, and Su. And the doomers are organizing the latest RMN “don’t kill us” party. The bazaar of ideas is bustling, blind prophets leading multitudes, psyops in every direction and the buyside analysts are updating their models accordingly.

[^16]: Noga Alon, Thomas F. Bloom, W. T. Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood, “Remarks on the Disproof of the Unit Distance Conjecture,” arXiv:2605.20695, May 20, 2026, 1–3. Tony Feng et al., “Towards Autonomous Mathematics Research,” arXiv:2602.10177, revised March 6, 2026, 1–2, 7–10.

The dollar-weighted opinion is unquestionably bullish. McKinsey dutifully projects $7T in data center spend through 2030,[^17] and that’s probably an underestimate. The smart money is rushing into labs, chips, neoclouds, and behind-the-meter power, while an ever-increasing torrent of private capital flows into AI. The duller ones say they’re more interested in *picks and shovels* and write checks to build tools for Anthropic researchers which are inevitably acquihired with little fanfare but much economic return within eighteen months. And behind this vanguard lies the deep end of capital markets: insurers, non-bank lenders, and the wealth-channel-packaged retirements of America’s doctors and dentists financing the credit side of the greatest infrastructure buildout ever seen this side of the railroads.[^18] These walls form the arena of the AI debate, the positioning and counter-positioning by which competing investors form their identities and views of the future. But because the plurality of capital moves *inside this arena,* its participants have mistaken its walls for the horizon. There is so much more to see.

[^17]: Jesse Noffsinger, Mark Patel, and Pankaj Sachdeva, with Arjita Bhan, Haley Chang, and Maria Goodpaster, “The Cost of Compute: A $7 Trillion Race to Scale Data Centers,” *McKinsey Quarterly*, April 28, 2025.

[^18]: Laura Benitez, “Apollo Muscles In on Wall Street’s Turf with $300 Billion of Loans,” *Bloomberg*, July 14, 2026, 1–2. Apollo Global Management, “Apollo Leads $35 Billion Capital Solution for Broadcom AI XPV Platform in Partnership with Blackstone and Leading Global Banks,” June 9, 2026. CoreWeave, “CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-Backed Financing,” March 31, 2026.

We must understand that the terms of debate describe only the *adolescence* of artificial intelligence, a coltish growth spurt rewarded with gobs of money. Before it was, work waited for it. AI was born only recently, and in San Francisco the model still stands in its **GENESIS.**

The Genesis is composed of everything inside the cradle, the factory required to design, train, evaluate, host, and serve a model. We managed to write the algorithm. Then we trained it. And now we’re hosting this artifact for a price that is hurtling downward toward the requisite compute.[^19] This will go down in history as one of the greatest human achievements of all time, but so far it still only amounts to the token generation loop.

[^19]: Luke Emberson and David Roodman, “The Plunging Price of Thought,” Epoch AI, September 22, 2026.

So far, its movement down into the labor process, beyond simple access to a free chatbot, has been minuscule. Few see how an API endpoint can augment their work, and fewer still have figured out how to economically benefit from it.[^20] It turns out that one of the most expensive artifacts wrought by man’s hand, or at least his design, produces impeccable decks, optimizes workflows exceedingly well, and files exceedingly detailed customer service complaints. The early adopters keep paying, increasing their usage, and giving the labs their first real revenue. For machine intelligence, this is just a summer job.

[^20]: Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger, and Aditya Pande, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” CES Working Paper 26-25 (U.S. Census Bureau, Center for Economic Studies, April 2026), 1–4, 23–27, 33–34, 58–59. Adam Grundy, Cory Breaux, and Dhanapati Khatiwoda, “Large Firms With at Least 20 Employees Biggest AI Users,” U.S. Census Bureau, May 26, 2026.

The man in Bangalore is not following Hacker News discourse. The small contractors, the middle-aged, the service workers and gig workers countless are laboring in a fog of ignorance fueled by underpowered Microsoft Copilot workflows. They are not reading DeepSeek API caching docs or thinking critically about their workflow or tracking the bleeding edge and deploying it into their business. The labs, relying just upon the vanguard of early adopters, have gotten the Genesis to world-historic levels of annualized revenue with little more than intelligence alone.[^21]

[^21]: “OpenAI Expects $70 Billion in Annualized Revenue by End of 2026,” *Bloomberg*, October 9, 2026. “Anthropic’s Quarterly Revenue Jumped 45% to Nearly $17 Billion,” *Wall Street Journal*, October 14, 2026.

On the other side of the railroads: the earliest productive use of locomotives was not for the intercity transport of people and goods. Instead, they were financed by ironmasters and colliery owners, who put them to work on captive industrial routes, where the traffic already existed, and a machine could do it better, faster, and cheaper.[^22] In 1804, at Penydarren, Trevithick’s early engine was installed on a road previously traveled by horses, and used to haul tons of iron nearly ten miles from the ironworks to a wharf, where it could be transferred onto a boat and shipped to Cardiff.[^23] Beginning in 1812, the first commercially viable locomotives replaced horses on the Middleton wagonway, tasked with pulling coal from the colliery at a lower cost.[^24] These were repetitive movements of heavy materials, directed by owners who already had a profitable horse-based operating system in place. They commissioned the locomotive builders to augment and improve an existing operation with the new technology of the era.

[^22]: David Gwyn, *The Coming of the Railway: A New Global History, 1750–1850* (New Haven, CT: Yale University Press, 2023), 1–2, 20–21, 92–100.

[^23]: Richard Trevithick to Davies Giddy, February 22, 1804, reprinted in Francis Trevithick, *Life of Richard Trevithick, with an Account of His Inventions*, vol. 1 (London: E. & F. N. Spon, 1872), 161–62. Jennifer Protheroe-Jones, “Birth of the Railway Locomotive,” Amgueddfa Cymru-Museum Wales, May 16, 2020.

[^24]: Gwyn, *Coming of the Railway*, 98–100. “Our History,” Middleton Railway, accessed August 11, 2026.

![Hand-coloured print of a collier with a pipe, a cane and a basket. Behind him Blenkinsop's steam locomotive hauls a train of coal wagons along the Middleton Railway, below a pithead with its winding gear and chimney.](https://commanding-heights.com/img/walker-the-collier-1813.jpg)

*George Walker, The Collier, 1813. Blenkinsop’s locomotive hauling coal on the Middleton-Leeds Railway.*

Picture credit: George Walker, *The Collier*, engraved by Robert and Daniel Havell (Leeds: Robinson & Son, August 1, 1813), aquatint, plate 3 in *The Costume of Yorkshire* (1814), Science Museum Group, object no. 1935-505.

This is the analog to the revenue that Genesis artifacts earn today: these early contracts gave the locomotive builders and the AI labs alike a reason to continue to develop their technology and production methods. But as we know, the lion’s share of the economic gains of the railroad age came not from building locomotives or integrating them into captive industrial routes, but from the network they made possible.[^25]

[^25]: Richard Hornbeck and Martin Rotemberg, “Growth Off the Rails: Aggregate Productivity Growth in Distorted Economies,” *Journal of Political Economy* 132, no. 11 (November 2024): 3547–3602, esp. 3550, 3583–85, 3592–93. The railroad sector captured only 7 percent of the railroads’ 1890 social return.

The step-change came when the locomotive moved off of the captive route that its owners had initially intended. The railway, now publicly accessible, was a boon for the surrounding economy. The openness allowed businesses and people to move freight themselves. Rail began to connect independent actors and made what was once disparate unified. The railroad network facilitated emergent economic activity that no railway owner had designed in advance. A new economy of scale unanticipated.[^26]

[^26]: Alexander Klein and Peter Matthews, “Transportation Costs in the Antebellum USA: A New County-Level Dataset with Time-, Region-, and Direction-Specific Freight Rates, 1820–1860,” *European Review of Economic History*, advance online publication, April 8, 2026, 2–5. Hornbeck and Rotemberg, “Growth Off the Rails,” 3547–51, 3582–85, 3592–93.

![Coloured print in two bands showing traffic on the Liverpool and Manchester Railway in 1831: above, a locomotive hauls wagons of goods under tarpaulins; below, a locomotive hauls open cars of cattle tended by drovers.](https://commanding-heights.com/img/liverpool-manchester-railway-1831.jpg)

*Traveling on the Liverpool and Manchester Railway, 1831. Published 1894.*

*Goods, cattle, and mail, waybills paid by ordinary economic actors, not railroad owners.*

Picture credit: *Traveling on the Liverpool and Manchester Railway, 1831* (London: Raphael Tuck & Sons, 1894), lithograph, The Henry Ford, Dearborn, Michigan, object no. 89.0.332.288.

Intelligence is still in the captive route. Revenue is driven by sophisticated customers bringing AI into an already existing workflow and reaping the rewards of the model’s new capabilities and cost structures. It takes sophistication to draft the brief. Clients must accurately handicap what work can be properly substituted, monitor the frontier for the latest Pareto model, and create high technology (bespoke .MDs) to instruct the models that still cannot continuously learn. Though soon they will. Managers must learn the spiky distribution of their new electronic employees, who talk just like people, but don’t act at all like the personages they’ve replaced. These requirements exclude all but a narrow population of enthusiasts and bleeding-edge firms.[^27]

[^27]: Among OpenAI’s enterprise customers, the top 10 percent of firms by output tokens per active user generated 8.3 times as many as the middle 10 percent in June 2026, up from 2.6 times in January. OpenAI, “Enterprise Signals: What Frontier Firms Are Doing Differently,” updated August 12, 2026.

It’s easier to hire people to force the adoption of technology. OpenAI’s own DeployCo and many others like it will send engineers to embed within your organization to do the diffusing for you.[^28] Where there’s a will, there’s capital. Services roll-ups become *AI* services roll-ups. They’ll buy the olden business to use as a training environment for the model, giving them the access to the ontology and authority over workflow that it’ll take to make it work. When the other bidders for your Midwestern accountancies have not yet learned how to deploy LLMs into the business, you can say the biggest number during the auctions and still expect to hit your underwritten IRR thanks to a small cohort of forward-deployed, stimulant-addled workflow engineers.

[^28]: OpenAI, “OpenAI Launches the OpenAI Deployment Company to Help Businesses Build Around Intelligence,” May 11, 2026, under “Building for Where Frontier AI Is Headed.” “OpenAI Creates New Unit with $4 Billion Investment to Aid Corporate AI Push,” *Reuters*, May 11, 2026.

Some of these attempts may very well become durable downstream businesses that make plenty of net income for decades to come, but their scope is constrained and far too small. The DeployCos are limited in their ability to diffuse AI to the world by the number of capable experts they can hire, and the roll-up platforms must similarly struggle to acquire and integrate each new add-on. Both are too slow to diffuse AI at the speed at which intelligence is becoming abundant. And when viewed relative to the opportunity set, their capital and headcount are so small they disappear from the map.

Investors, meanwhile, build their understanding from metrics familiar to most: usage, OOMs, hockey sticks, ARR, and reading credit card data like the flight of the birds in Roman times. When you see the greatest numbers you have ever seen for any product past or present, it should be obvious where to invest. Then you look to your left and your right and see all the sharps investing there too. Slowly, the blinders form to block anything but those metrics, and situational awareness, however robust its owner believes it to be, contracts around a certain San Franciscan summer of love.

Be merry. There is plenty to celebrate, now. Lab ARR is extraordinary even if slightly ailing. American inference margins are, surprisingly, positive for now. Gross profit is being booked by CFOs,[^29] and investors are popping champagne over generationally large post-money valuations and impending IPOs.[^30] These metrics prove that Genesis is Something People Want, and people are paying for it. But a larger bill is yet due. The eventual intelligence economy will be orders of magnitude larger than Genesis, even Genesis fully matured.

[^29]: Sri Muppidi, “Why OpenAI, Anthropic Missed Their Own Gross Margin Forecasts,” *The Information*, February 24, 2026. Cory Weinberg, “Anthropic Reports $16.8 billion in Third-Quarter Revenue,” *The Information*, October 14, 2026. The reported margins are after inference costs but before research, sales, and other operating expenses. More recently, Anthropic reported an adjusted operating margin of 11 percent for the year through the third quarter of 2026, excluding stock-based compensation, and positive cash flow in the third quarter. “Anthropic Tells Investors Quarterly Revenue Up Nearly 50% to $16.8bn Ahead of IPO,” *Financial Times*, October 14, 2026.

[^30]: Echo Wang, “Anthropic IPO Stands Alone as AI Market Jitters Sideline Others,” *Reuters*, October 22, 2026. Shirin Ghaffary and Rebecca Torrence, “OpenAI Shifts Value to $1.25 Trillion as Funding Talks Finalize,” *Bloomberg*, October 23, 2026.

As the newly capitalized ARR comes in, it is immediately put to use. Inside the labs, the cash supports even larger training runs still. Outside of the labs, the revenue run rate is capitalized at stratospheric multiples,[^31] summoning forth new companies and new pools of capital toward every new trade that promises to keep earning rents.

[^31]: Anthropic, “Anthropic Raises $65B in Series H Funding at $965B Post-Money Valuation,” May 28, 2026. OpenAI, “OpenAI Raises $122 Billion to Accelerate the Next Phase of AI,” March 31, 2026. Jaspreet Singh and Krystal Hu, “OpenAI’s September Annualized Revenue Nears $50 Billion, Less Than Previously Indicated, Source Says,” *Reuters*, October 8, 2026, updated October 9, 2026. The March and May completed-round valuations correspond to approximately 17 times OpenAI’s September annualized revenue and less than 14.9 times Anthropic’s July annualized revenue, respectively.

Historically, the fashionable and consensus businesses that investors have fawned over as possessing the ultimate attributes have thrown off more free cash flow than their operating model can absorb in reinvestment at attractive rates of return. Genesis is nowhere near that point.

The labs will continue to devour the cash flows of the hyperscalers, and whatever dollars the capital markets wish to supply.[^32] There will always be dollars available. We expect Genesis to grow prodigiously throughout this process.[^33] Superusers will consume more and more productively, their hunger unabating. And there will be a growing class of users who will be sought and transformed by intelligence reaching through a digital network. The latest marginal user arrives, then the next one. Billions to come.

[^32]: OpenAI, “OpenAI Raises $122 Billion to Accelerate the Next Phase of AI,” March 31, 2026, under “Deep Conviction across Global Capital” and “Compute Is a Strategic Advantage.” Patturaja Murugaboopathy and Gaurav Dogra, “AI Investment Boom Puts Big Tech’s Free Cash Flow under Pressure,” *Reuters*, July 22, 2026. Tatiana Bautzer and Gertrude Chavez-Dreyfuss, “Banks Get Creative and Look Further Afield as AI-Fueled Debt Soars,” *Reuters*, June 29, 2026. NVIDIA, “NVIDIA Partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize over $500 Billion of Third-Party Capital,” August 10, 2026.

[^33]: OpenAI, “Introducing GPT-6.1 Sol,” September 29, 2026. Anthropic, “Introducing Claude Haiku 5.5,” October 7, 2026.

And there will always be a binding constraint du jour. The velocity of ideas from certain Signal group chats to discourse on Twitter to local daytime TV to White House executive orders is increasingly high. They’re saying words in the Senate that have never been said before. First GPUs, CPUs, SSDs, HBM, GWs, MTok/s, and then the next concepts which have not yet become buzzy initialisms. As one demand is filled and margins fall, the next in line becomes capital’s target. That scale may itself enable Genesis to grow still larger, richer, and more central for decades yet. But what could cause this buildout to stop?

Normally, this problem is solved by the capital cycle. Excess returns incentivize entry and new capacity, which eventually outstrips demand, resulting in falling utilization and incremental returns. Once the incremental returns fall below the market rate of return on capital, new projects fail to clear their hurdle rates, capex slows, the tourists get washed out, and the industry eventually recovers. This is capital’s way of allocating a scarce resource, in this case, itself.

But in Genesis, there may never be a signal that tells participants to stop building. You can’t simply ask what the return on capital is for a new incremental build of a laboratory and underwrite it that way, because if you fall behind the frontier, you risk ceding all of the value that you’ve previously invested in. It is a common fiction that each model generation is a new hand of cards; the intelligence trade is one persistent and everlasting parlay on a civilizational scale. And the hyperscalers aren’t investing in one model or even one data center cluster; they are buying call options on future profit pools that they can use to defend the profit pools that they already have. Their calculation is made over the breadth of their business, and the cost of not investing is informed by what their rivals might do.

This means that, even as the returns on incremental investment in Genesis come down, it may not be irrational for participants to keep investing, even if they acknowledge that the industry is overbuilding. **In other words, what is rational capital allocation for a particular firm may be capital destruction in the aggregate.** The ouroboros is no longer fed by ROIC, but by the coercive laws of competition.

As frontier models continue to acquire military, cyber, and intelligence value, the rational response to additional training runs shifts again. Any technology at a level of creative and destructive capability that implicates the whole of any polity, or perhaps the world, naturally becomes a matter of preoccupation for the sovereign. It was known first at Trinity, and every moment since then has been a contest to own the most capable nuclear deterrent, and to prevent your rivals from doing the same. Immediately after Hiroshima and Nagasaki, it seemed possible that a unified, global sovereign would assume control over the dispensation of atomic technology. But at the end of the twentieth century, this question was resolved by proliferation in favor of diffused access to a MAD stalemate.

The same contest is once again at hand at the frontier of machine intelligence. Capabilities will very quickly enter the indistinct realm of sovereign combat, cyberwarfare, and military projection. The increasing resources allocated to train the frontier will then start to look less like a question of capital reinvestment, and more like the construction of a carrier strike group: A strategic asset operated for the capability it confers. In private and public conversation, the leaders of all of America’s core intelligence functions report even current capabilities decreasing time to vulnerability by orders of magnitude, even in our current sclerotic intelligence apparatus. It is by these metrics that the development of machine intelligence will be seen in Beijing and Washington. And by whom will the capability be wielded? Whatever comes to pass, the interest of the sovereign will play a final, decisive role.

It is clear that Genesis is going to succeed, if success is meant to be understood as continuing. Whatever happens to its private returns, the frontier is advancing and will continue to do so. The capital calls are being met, the turbines are coming online, tokens are becoming more abundant and cheaper, the models producing them, more capable. More, more, more. The lab as factory. The loading dock is never quiet. The sign is up, painted in large letters somehow visible in all languages at once: Intelligence for sale.

No thing that relies upon model pricing margins will escape the walls of Genesis. But where will the balance of value settle, and to whom will it accrue? That is the question of the **COMMANDING HEIGHTS.**