The buildout behind artificial intelligence has real strengths: model-maker revenues are compounding, and the deepest-pocketed balance sheets are backing many transactions. The open questions are chip resale values and the pace of new silicon generations. The decisive variable is durable corporate demand, and on this the jury is out. Companies with strong growing recurring customers rarely default and so when you meet a bear on these financing structures, ask when they last spoke to a corporate about long-term AI adoption, because that is where the true risk lies.

Subscribe now

Nvidia’s reported willingness to guarantee $250 billion of financing for an OpenAI data center produced the sharpest question of this cycle, posed by Jim Chanos in late July:

Why does the seller need to provide financing guarantees for roughly two-thirds of the cost of the chips it is selling?

Credit markets asked the same thing, and Nvidia’s default swaps recorded their largest intraday jump since the contracts began actively trading. The practice being condemned has a name, vendor financing, and the condemnation treats it as a late-cycle trick, something companies resort to when real demand dries up.

Suppliers have lent customers money to buy their products since before the Civil War. The record contains spectacular successes, catastrophic failures, and enough of both to extract the pattern that separates them. That pattern, not the word circular, is the right instrument for judging the buildout now assembled behind graphics processing units, the GPUs on which artificial intelligence runs, because no vendor-financing program of this size has ever been attempted.

The structure was invented to solve a timing problem that banks refused to touch. In the 1840s and 1850s, Cyrus McCormick sold mechanical reapers to farmers who had no cash until the crop came in, so he let them pay across harvest seasons, telling buyers that “it is better that I should wait for the money than that you should wait for the machine that you need”. By 1860, nearly 70% of western farmers who adopted the technology were using his. Isaac Singer’s partner Edward Clark went further in September 1856 with the first installment plan run at national scale, offering a $110 sewing machine for $5 a month at a time when a typical family earned about $500 a year. Sales tripled within 12 months, and by 1880 Singer was producing more than half a million machines annually. The customer’s cash arrived later than the customer’s need, and the seller understood the machine’s value in a way no bank could.

The railroads then added the piece that still governs aircraft finance today. In March 1842, the Locks and Canals machine shop agreed to build running gear for 600 coal cars for the Philadelphia and Reading Railway Company, with title held by two trustees until the railroad had paid in full and payment reserved at 22.5 cents per ton of coal hauled. One year later, the same parties financed twelve locomotives and 450 more cars on two-year terms. That contract, nearly two centuries old, contains two inventions that survive in every equipment-finance contract written since: title stayed with the financier, so a railroad bankruptcy meant the cars came back rather than joining the estate, and payment was tied to tons hauled, which is power-by-the-hour economics more than a century before Rolls-Royce made the phrase famous.

Philadelphia’s trust companies standardized the structure into equipment trust certificates issued to about 80% of the equipment’s value and amortized roughly 10% a year, and that instrument, carried from railcars into aviation, became the modern enhanced equipment trust certificate, the EETC of aircraft finance, in 1994. The manufacturer’s motive in 1842 is equally recognizable, because the shop’s directors backed the deal “to give profitable employment of our shop and foundry at a time when we have little else to do”. Vendors have been financing demand forward in slack markets from the very beginning.

The automobile turned the practice into an institution. Banks would not finance car purchases, a refusal Alfred Sloan, writing in his 1963 memoir, attributed to bankers who “thought of the automobile as a sport and a pleasure, and not as the greatest revolution in transportation since the railway”. Seeing this market gap, General Motors founded the General Motors Acceptance Corporation, GMAC, in 1919 and financed its own customers. By 1930, three of every four cars and trucks in America were bought on credit, and Henry Ford, who considered consumer credit morally reprehensible and asked buyers to save up through their dealers, watched his dominance collapse against a competitor that was prepared to fund today’s purchase. Ford eventually surrendered and incorporated Ford Motor Credit in 1959; it carries $158 billion of assets today.

Academics have only recently formalized why the captive works: a manufacturer that finances its own product commits to supporting that product’s resale value, which improves collateral for every lender in the market. A century of experience matches the theory, because GMAC carried car buyers through depressions, wars, and oil shocks without incident, and its one near-death experience, in 2008, came from its mortgage subsidiary rather than from auto lending.

Aviation – A Proven Model

Aviation is where the structure reached maturity. A three-firm oligopoly builds aircraft engines, with CFM International, the General Electric and Safran joint venture, powering 72% of the in-service narrowbody fleet, while GE and Rolls-Royce split widebodies. Rolls-Royce sells widebody engines at a loss, £1.1 million to £1.6 million per engine on its own disclosures for 2017 through 2020, and recovers everything through TotalCare contracts priced per engine flying hour over decades, the railway car leasing per-ton payment of the 19th century carried into the jet age. Services now produce 69% of its Civil Aerospace revenue. Nobody paid cash for any of it: leasing grew from roughly 10% of the world fleet in the 1970s to 58% by the end of 2023, and the engine makers sat inside the financing loop themselves, with Rolls-Royce co-owning a spare-engine lessor holding more than 900 engines and Safran running the world’s largest lessor of CFM engines. The financiers wrote enormous checks against a concentrated supplier base and rented the capacity onward to hundreds of airlines, holding the residual risk in between.

Airlines failed constantly across the period, so the aviation leasing loss numbers are not evidence of default per se; they reflect recovery, and recovery is a property of the resale market rather than the borrower. Aircraft-backed certificates produced a cumulative loss-given-default of 0.89% on senior tranches across three decades, and the median bank loan secured by aircraft recovered in full, because a repossessed aircraft re-lets into a deep, liquid, global market of operators within months. When the COVID-19 pandemic erased 98% of passenger demand, Moody’s held the ratings of the seven lessors it reviewed. AerCap lost $299 million in its worst pandemic year with fleet utilization at 97%. Two years after Russia confiscated more than 100 of its aircraft, AerCap was selling planes at $651 million above their book value. Vendor financing at its best is not a bet on the borrower at all. It bets on the asset’s next user.

Itel – Failure on a Grand Scale

The computer-leasing collapse of 1981 is the GPU model’s direct ancestor. IBM historically rented rather than sold its machines until a 1956 antitrust consent decree forced it to sell them, creating a secondhand market and opening the door for third-party lessors. Itel was the most aggressive, founded in 1967, buying IBM mainframes with 20% equity and 80% debt and leasing them out below IBM’s rates. Within twelve years it had written $1.7 billion of leases, second only to IBM itself. Then, in 1979, IBM launched the 4300 series, faster and up to 30% cheaper than the machines on Itel’s books, and the residual values supporting Itel’s structure evaporated with a competitor’s product announcement. Itel lost $433 million in 1979 and filed for Chapter 11 in January 1981, owing $1.3 billion. The lessor’s collateral was tied to the manufacturer’s product cycle, and the party bearing the residual risk had no control over the event that destroyed it. IBM, having watched the lessors fail, set up its own captive finance arm that same year.

Telecom – The Low Point

Telecom’s version of vendor financing, two decades later, failed on every front. The US Telecommunications Act of 1996 opened America’s local phone markets to competition, and equipment vendors financed the 711 startup carriers that rushed in, most of them with nothing more than a business plan. Lucent had committed roughly $7 billion of credit by the end of fiscal 2000, and the lending had stopped pretending to be equipment finance: Lucent funded more than 100% of one customer’s equipment purchases, and Nortel went as high as 130%, which means the vendors were paying their customers’ payroll and marketing to keep the orders flowing. The switches had no second buyer; demand never arrived, with only 2.7% fiber utilization by 2002; and the borrowers had nothing underneath them, as Winstar showed when it collapsed owing $5 billion against $200 million in assets. The US Securities and Exchange Commission later found $1.148 billion of fraudulently recognized revenue inside Lucent’s fiscal 2000 results. Lucent lost $16 billion in fiscal 2001 and disappeared into Alcatel, and Nortel eventually liquidated all legacy assets. Worthless collateral, absent demand, insolvent borrowers, and fraud layered on top: every failure the history warns about, packed into a single episode.

Why Some Vendor Financing Fails

The history of vendor financing shows numerous successes and at times, dramatic losses. Vendor financing never failed because capital moved in a circle, since that circle was present in every success, from the Reading coal cars to Rolls-Royce’s captive lessor. It failed for three reasons.

1. Collateral: cars and aircraft recovered because deep resale markets stood behind them, while Itel’s mainframes were destroyed by the manufacturer’s product cycle and Lucent’s switches had no alternative user at all. The depth and liquidity of the secondary market are key leading indicators of recovery rates in the event of default.

2. Demand: farmers harvested, drivers drove, and passengers flew, while the bandwidth demand of 1999 was a forecast that missed by years. If customers keep buying the underlying products, defaults are rare, and resale value is robust. Manufacturers with a strong customer base rarely default.

3. Residual risk: the captives that financed their own product survived everything, while the third parties that bore residual risk they could not control were periodically wiped out. New product cycles can undercut the value of equipment already financed, and a third-party lessor has no control over that event, while the manufacturer that finances its own product is committed to defending its value.

Defaults were commonplace across many sectors. The strength of the underlying model depended on what a lender recovers when a borrower fails, whether the demand behind the asset is real, and who ends up holding the loss if the asset’s value collapses.

GPU Vendor Financing: Next Level

The scale of the GPU financing model is unprecedented. Hyperscaler capital spending ran near $380 billion last year and is guided toward $725 billion by the end of 2026. The funding has moved into public markets, where five technology companies issued $159 billion of bonds in six months, and data-center securitizations have grown fifteen-fold in six years to $61 billion. Nvidia has announced a $500 billion financing platform with six Wall Street firms to lend customers the money that returns as revenue, against a telecom episode that peaked near $25 to $30 billion in total. Every structure from history is in use: equity stakes such as Nvidia’s $30 billion in OpenAI, take-or-pay capacity contracts as large as Oracle’s $300 billion, GPU-collateralized debt rated investment grade, and off-balance-sheet trusts, led by Meta’s Hyperion vehicle with Blue Owl, that a Philadelphia trustee of 1880 would recognize on sight.

The concentration is familiar too, since two customers have supplied as much as 39% of Nvidia’s revenue in a single quarter, a tighter funnel than the engine makers ever ran.

History’s loudest warning to this structure concerns the collateral, and the resale market for accelerated computing is growing but untested. H100 rental prices fell by roughly a third from their 2024 peak, and while six-year-old A100S still clear at about $19,000 and remain under contract, no default cycle has ever forced GPUs onto the market at scale. For custom silicon such as Google’s tensor processing units, its TPUs, there is no resale market at all, because the chips are rented inside one vendor’s cloud and never sold.

The industry cannot even agree on what the asset is worth, since Amazon shortened server depreciation lives to five years while Meta lengthened them to five and a half. Underneath everything sits the Itel precedent, because the event that destroyed computer leasing was not a recession but the manufacturer’s own next-generation product, priced 30% cheaper. Nvidia now ships a new generation on an annual cadence, which means the IBM 4300 mainframe launch equivalent happens every year, and the parties bearing the residual risk of last year’s chips are mostly not the manufacturer.

Aviation never faced the second vulnerability: the funnel in the middle of the chain. AerCap spreads its book across roughly 300 airline customers, while CoreWeave took about two-thirds of its 2025 revenue from Microsoft alone, and the broad corporate demand for inference reaches the financiers through a handful of labs. Breadth at the end of the chain protects the middle only if idled capacity can be re-let to a substitute tenant, and there are three to five candidates rather than three hundred.

The collateral is also correlated with the counterparty in a way no car or aircraft ever was, because the scenario that defaults a lab is the same scenario that craters GPU utilization and resale prices. The commitment ratios have no historical parallel either, since OpenAI has reportedly committed on the order of $750 billion of compute spending through 2030 against current revenue near $2 billion a month, and no airline signed leases at twenty times its revenue. What history counts in the model’s favor is the seat of residual risk, because the hyperscalers are captives in the GMAC sense, consuming their own capacity through their own products, and the credit market prices the distinction exactly as the pattern predicts: Meta borrows at barely 50 basis points over Treasuries while CoreWeave pays 550 basis points over SOFR, the secured overnight financing rate, with covenants attached and Oracle’s default insurance sits at record levels.

The lesson of this history is that vendor financing was never really about the size and scale of the leverage. It was about the value of the collateral underneath, which is why railcars and aircraft built century-long records while mainframes and fiber switches ruined their financiers. For the AI industry and the sustainability of its chip-financing structures, everything reduces to three questions.

1. Will end demand from corporate adoption of AI tools grow fast enough to keep OpenAI and Anthropic ahead of the commitments they have signed?

2. Are we making too many chips against forward demand and risking a glut that crushes collateral value?

3. And will innovation itself sabotage the collateral, with each new chip generation doing to today’s GPUs what IBM’s 4300 series did to Itel’s mainframes in 1981?

We certainly do not have the insights to answer any of the three today, and bulls and bears alike are relying on incomplete information when determining a forward strategy. We optimists believe the early evidence leans constructive, with Anthropic’s revenue run-rate moving from $9 billion to $47 billion in roughly five months and tokens processed at Google up 330-fold in two years. Those with concerns point to the lack of current ROI for corporates as a sign of a pending retraction. Both arguments have flaws, and the only truth is that it is too early to tell.

Ultimately, all three questions collapse into one variable: long-term corporate demand. Without it, there are few good outcomes for the AI ecosystem or the assets underpinning this transformation, however cleverly the paper is structured. The bears on these financing structures are therefore looking in the wrong place. Financial engineering is not the fault line. Demand, along with its trajectory over the coming decade and companies’ willingness to redesign how they operate around artificial intelligence, will determine whether the whole edifice holds. So when you speak to a bear on AI vendor financing, ask a simple question:

Have you spoken to a corporate about its long-term AI adoption strategy?

That is where the true risk lies.

Subscribe now

The loudest worry about the AI buildout is that its financing is circular and therefore fragile, with chipmakers funding the labs that fund the clouds that buy the chips. The Google, Broadcom, and Anthropic transaction is the most advanced example of that machinery, and read carefully, it says close to the opposite of what the alarm assumes. It is not a hidden weakness dressed up as innovation. It is a piece of financial engineering that takes a difficult funding problem, the need to bankroll tens of billions of dollars of fast-moving hardware for a young and unprofitable tenant, and solves it by distributing the risk across several parties in the order of their ability to bear it. What the structure tells us about AI fragility is that the brittleness is not where the circular-financing story looks for it. The financing layer is stronger than it appears, and the real weakness is smaller, sitting in a narrow tail and in the cruder deals that lack this design.

How the structure works

Six parties sit in this financing:

Google. The chip vendor and sponsor. It designs the tensor processing units (TPUs), sells them into the vehicle, keeps Anthropic as a cloud customer, and stands behind the asset as the natural buyer of last resort because it needs these chips for its own workloads.

Anthropic. The tenant. It leases the chips under a multi-year, take-or-pay contract and pays the rent that services the debt, without owning the hardware or carrying the borrowing itself.

Compute SPV. The special purpose vehicle. A standalone company that exists only to own the chips and the debt raised against them, buying the TPUs from Google and leasing them to Anthropic.

Morgan Stanley. The arranger. It structured the vehicle and placed the debt with investors.

Private credit investors. The lenders. Apollo, Blackstone and other institutions supply the senior debt and the junior notes that fund the chip purchase, and are repaid from Anthropic’s rent in order of seniority.

Broadcom. The guarantor. It co-designs the TPUs and provides residual value support on the senior debt, agreeing to cover any shortfall if Anthropic stops paying and the repossessed chips sell for less than the balance owed.

Subscribe now

According to reporting in The Financial Times, Google sits at the center of roughly $200 billion in financing arrangements for Anthropic’s infrastructure, of which more than $150 billion is Google’s own TPUs. The transaction is easiest to follow as a sequence. Google sells about $35 billion of its TPUs into a special purpose vehicle called Compute SPV, a standalone company created only to own these chips and the debt raised against them, which moves the borrowing off Google’s balance sheet while Google keeps Anthropic as a customer. Morgan Stanley is the banker on the deal, and the vehicle pays Google for the chips using money raised from outside investors: roughly $30 billion of senior debt anchored by Apollo and Blackstone, a $4.5 billion tranche of lower-ranking B-rated bonds, and a thin sliver of equity beneath both. The vehicle then leases the chips to Anthropic, and that rent is the vehicle’s only real income. The rent flows back through the vehicle to pay interest and principal to the investors in order of seniority, and Broadcom guarantees roughly $30 billion of the residual, covering the shortfall if Anthropic stops paying and the repossessed chips sell for less than the debt outstanding.

The result is a clean division of responsibilities: Google offloads the hardware and the financing risk while keeping the customer, the investors hold paper backed by Anthropic’s rent and Broadcom’s guarantee, Anthropic runs a million chips it neither owns nor borrowed against, and each party carries the slice of risk it is best placed to price. The template is borrowed from aviation, where Boeing and GE used similar vehicles for many years to finance aircraft and engines.

Why the design is strong

The intelligence of the structure is in its layering, and first-loss provisioning is the heart of it. Below the senior debt sits a thin equity layer and a $4.5 billion tranche of B-rated bonds that carry no residual guarantee, and these absorb the initial losses before any senior lender is touched. S&P flagged that unguaranteed tranche as credit negative, which is not a flaw but the system working as intended, because the market is pricing the junior risk separately and the investors who buy it are paid to stand in front. Senior capital, the bulk of the money, is insulated by everything stacked beneath it, and it is shielded again by Broadcom’s guarantee behind it. That double protection is why the paper prices where it does. The senior debt split into a $6 billion bank tranche at 100 basis points over Treasuries and a $24 billion institutional tranche near 5.75%, both privately rated in the mid-investment-grade range, while the $4.5 billion junior notes priced around 8.5% to reflect their unguaranteed position.

Across the wider market, the same backing lets TPU operators borrow near 7.1% while comparable Nvidia-ecosystem borrowers pay about 9.3%, a two-point advantage created by Broadcom’s support and the first-loss provisioning. The design converts a hard-to-finance asset into investment-grade paper by allocating the difficult parts of the risk to the parties that understand them, and that is a strength rather than smoke and mirrors.

Scenario testing the transaction

The scenario people reach for is a default, and the base case comes first because it is the expectation rather than the tail. Given that the lease is take-or-pay, Anthropic owes the rent regardless of use, and if that rent amortizes the senior debt over the life of the lease, it repays the lenders in full with no reference to what the chips are worth at the end. In that base case, the leftover chip value is upside for the equity holders, and Broadcom’s guarantee is not how the lenders normally get repaid.

It does two other things: it wins the deal a higher rating and cheaper funding, and it protects the lenders only in an early default, when the lease breaks before the loan is paid down. Only in that tail does the resale value of the chips matter at all. The table and chart below run it from a 95% recovery down to a distressed 50, and it is the day-one picture, before any principal has amortized.

Losses land in a strict order on the parties who signed up for them, and the senior lenders stay whole across every case shown. The first-loss equity goes first, then the B-rated tranche, and only once recovery falls below 80% does Broadcom’s guarantee begin to pay, exactly as the layering intends. The junior notes do not cover a fixed percentage of the deal. They cover a fixed dollar band, the shortfall between roughly $0.8 billion and $5.3 billion, wherever it falls. The modeled recovery is unlikely to lead to losses because used AI chips have been trading at 60% to 80% of their purchase price for older A100S, and H100S coming off expired 2022 contracts rebooked at about 95% of their original pricing. Re-leasing is a superior outcome to liquidation, since renting idle chips generates two to four times more total revenue than a one-time sale over an eighteen-to-twenty-four-month window, so a distressed dump is the worst path rather than the likely one.

Put the base case and the tail together, and a senior lender loses only if an early default, a distressed sale, and a reluctant backstop all arrive at once, which is a demanding set of conditions by design.

Remember, all this analysis assumes a Day 1 default.

The waterfall above holds the senior debt at its full $30 billion, which is the picture only on the day the vehicle is funded. As the rent covers interest and principal together, the senior balance falls every quarter, and the recovery needed to keep those lenders whole falls with it. On the reported 5.75% senior coupon over an illustrative four-year term, roughly $15.9 billion remains at the end of year two and about $8.2 billion at the end of year three, so the chips would have to sell for only 45% of their original cost to cover the year-two balance and 23% to cover year-three. Set that against a used market where these accelerators have resold at 60% to 95% of cost, and the senior lenders are covered with a widening margin from roughly the second year onward, well before the loan matures.

Amortization also takes most of the work away from Broadcom’s guarantee. On day one, a 50% sale would cost Broadcom up to $12.2 billion. Once the debt has been paid down, the same distressed sale at a year-two default is absorbed by the equity and the junior bonds, and Broadcom pays nothing. By a year-three default, every tranche is whole. The risk is therefore concentrated in the first year or so, when little principal has amortized, which is also the window in which the chips are newest and Google is most eager to take them back, so the two protections strengthen at the same moment. This works as long as the debt is paid down at least as fast as the chips lose value. A normal repayment schedule does that; an interest-only loan would not. The exact schedule is not public.

Broadcom bears the contract, Google stands behind the asset.

Broadcom carries the explicit residual obligation, covering roughly $30 billion if the chips come back and have to be sold, but Broadcom is a chip designer with no use for a million TPUs, so what makes its guarantee cheap to honor is the presence of a buyer ready to pay near the chips’ worth. That buyer is Google, and the point deserves emphasis because it is the load-bearing reason the structure holds. Google is short of its own TPUs, rationing them so tightly that its DeepMind researchers queue for capacity behind paying cloud customers, and it sold Anthropic silicon that Reuters described as capacity traditionally reserved for internal use, against a Google Cloud backlog of $514 billion as of mid-2026. Google is therefore not a hopeful reseller who might appear at an auction; it is a starved end user of these exact chips that would take them back to feed Gemini, Search, and Ads inference it cannot supply today.

In a default, the most probable outcome is not a distressed sale into a thin secondary market but Google reabsorbing a million of its own chips at close to fair value, which is why the residual on this deal is sturdier than a generic pool. The one honest qualifier is that Google’s role is an economic near-certainty rather than a signed purchase obligation, so this note treats Broadcom as the backstop of record and Google as the deep and highly probable buyer standing behind it.

Is a Google repurchase guaranteed? No, but it should give senior lenders significant comfort.

The chip’s second life

The residual is not a leap of faith, because the secondary market for these chips already exists. Chips leave the frontier fleet for density and power efficiency rather than failure, and they remain strong for inference and mid-size training, which is why used chips hold much of their value and turn over through dedicated resellers. The second life is not generic cloud, since documents and photographs are served on commodity processors and cheap disk, and a matrix-math accelerator instead cascades down a curve inside AI itself, from frontier training in the first two years to inference in years three and four to batch work thereafter. An older GPU/TPU runs production inference, then cost-sensitive batch inference where latency does not matter, then fine-tuning of smaller models, and inference is the fastest-growing of those workloads rather than a shrinking one. For the TPUs in this deal, the cascade ends inside Google, which needs exactly this capacity for its own inference and cannot get enough of it. The facility is durable on its own terms as well, since the shell, the cooling and above all the secured power outlast several chip generations, and a gigawatt of contracted power is among the scarcest assets in the industry.

Where the fragility remains

The structure has one true weakness, and it is correlation rather than design. The layers spread risk across equity, junior bonds, Broadcom and Google, and that spreading holds in almost every state of the world, but the parties re-converge in the tail. The scenario that pushes chip recovery toward the distressed end is a broad AI demand disappointment, and that same event weakens Anthropic as a tenant, pressures Broadcom as guarantor, and makes Google a more reluctant buyer at the exact moment its appetite is needed. The diversification works across normal conditions but is thinner in the one correlated downturn, which is the honest limit on how much comfort the structure should provide.

Whether that weakness stays small is a question of scale, not structure. On this deal alone, the losses are capped and the firms absorbing them are strong enough to take the hit. But the same setup is being repeated across AI finance: a guarantor whose own profits fall with the chips it guarantees, and a rescue buyer whose appetite disappears exactly when it’s needed most. Each deal looks manageable on its own. Stack up enough of them, all leaning on the same handful of strong hands and all exposed to the same downturn, and that is how systemic events happen, not one big failure, but many small ones arriving together.

Ultimately, it does not matter how well these securities are structured. The tranches, the guarantees, and the amortization schedules all sit downstream of a single question: can OpenAI and Anthropic keep driving demand? The customer base for this entire financing edifice concentrates on two companies, and their fortunes will determine not just the credit quality of these debt securities but the success of the AI rollout itself over the next five years. Their fortunes, in turn, rest on corporate adoption: on whether the models they train produce end products good enough that businesses pay for them at scale, year after year.

Structure can spread risk; it cannot create demand. A market this dependent on two customers is not a healthy one. But concentration is a vulnerability, not a verdict. If the products keep improving and adoption keeps broadening, these structures won’t be tested. The point is that everything rides on that “if,” and no amount of financial engineering can change it.

A weaker structure: CoreWeave and Nvidia

Not all transactions are as sturdy as Google/Broadcom/Anthropic. CoreWeave funds its buildout largely through GPU-backed delayed-draw term loans, carrying roughly $25 billion of debt across six such facilities, much of it resting on a small number of customers including an OpenAI contract worth $11.9 billion over five years. The backstop that supports this arrangement is not a tranched guarantee but a single order from Nvidia, disclosed in a CoreWeave filing, under which Nvidia is obligated to purchase the residual unsold capacity through April 13, 2032, an agreement with an initial value of $6.3 billion.

Three weaknesses separate it from the Compute SPV.

1. The backstop is capped at $6.3 billion against a debt load roughly four times larger, so it cushions rather than covers, and CoreWeave carries the balance of the exposure itself rather than an outside guarantor.

2. There is no layer of first-loss capital standing in front of the lenders the way the Compute SPV places equity and junior bonds ahead of its senior debt, so a single anchor customer that stops paying pushes the loss straight toward them.

3. And the backstop buyer is the chip vendor rather than an end user, so Nvidia is paying to keep demand for its own chips alive rather than consuming that capacity in a workload of its own, which recycles the demand question instead of answering it and lets the same dollar make several passes as Nvidia earns at manufacturing, at financing, and again at the backstop.

The Google structure spreads risk across parties and rests it finally on a captive user that needs the hardware; the CoreWeave structure concentrates risk on one customer, leaves the parent carrying the balance, and leans on a circular foundation from the company that sold the chips. That is why an AI disappointment would reach CoreWeave’s lenders long before it reached the Compute SPV’s, and the market already senses the difference, having repriced the whole complex without pushing it to distress, with Nvidia’s five-year credit default swap climbing to a record 82 basis points in late July from roughly 42 in June while still sitting far below the levels that signal a distressed borrower.

Why amortization makes even the weaker deals resilient

The amortization point goes past the guaranteed structure and changes how the whole complex should be read, including deals with no guarantor and no first-loss buffer. Every quarter of principal repayment lowers the balance an investor must recover, so time itself is a form of protection that builds automatically. On a four-year amortizing loan, the outstanding principal falls to a little over half by the end of year two, which drops the recovery needed to make senior lenders whole to roughly 45% to 50% of the chips’ original cost. That reframes the debate people are having about used-chip values. Whether a simultaneous flood of returned accelerators would clear above fifty cents on the dollar is a fair question, but it only bites on the earliest tranches if a default arrives almost immediately, because within two years the breakeven has already fallen to about that level and keeps dropping. The first tranches of these three-to-five-year structures are therefore the safe ones, not the risky ones, since they amortize into safety while the capacity that might one day pressure resale prices is still being built.

The real systemic risk is the customer, not the collateral, and it is concentrated rather than diffuse. OpenAI and Anthropic sit behind most of these leases, so a default on one contract signals distress that would likely surface across the others at the same time, which is the correlation that makes the tail real. That risk, though, is a near-term question with a clear horizon. If the two labs keep paying for the next twelve to eighteen months, the first cohorts of debt amortize into positions that survive even a harsh repossession, and the exposure that remains sits with later tranches financed against capacity that has not yet been built.

The CoreWeave and Oracle deals show the same clock at work despite lacking the Compute SPV’s guarantor and its first-loss layer. They carry more risk because a stressed customer hits their lenders directly, with no equity tranche and no Broadcom standing in front, yet they only get into real trouble if OpenAI or Anthropic default inside that same twelve-to-eighteen-month window, which on current funding and revenue trajectories looks unlikely. The cushion is cleanest in the true project-financed tranches that amortize on schedule, and it is weaker where a balloon payment or a corporate structure sits on top, so the resilience is a matter of degree rather than a blanket guarantee. Set the pieces together, and the conclusion is not that the complex is safe, but that its safest layers are the senior, amortizing, near-term ones, and its risk is a dated bet on two customers rather than an open-ended bet on the resale price of a chip.

To be clear, any default/reneging by the largest LLMs is a disaster for the AI industry. Even if it does happen in 18 to 24 months, when the amortization of the principal of the first generation of financing has reduced the amount owed to below 50% of the original face value, companies like CoreWeave face an existential hurdle because their ability to execute new higher-margin transactions plummets, and they would be left with decaying assets linked to older chips. While the “credit event” is taken off the table and the lender will likely recover in full (assuming a flood of chips can sell at 50 cents on the dollar), the CoreWeave business ceases to function, and it slowly dies rather than spectacularly blowing up.

Conclusion - What it tells us about AI fragility

The Google/Broadcom/Anthropic structure argues that the AI buildout is less fragile than the circular-financing story claims, and brittle in a more specific place than the conventional thinking believes. The innovation is that risk is not hidden but allocated, with first-loss capital in front, priced junior debt behind it, a manufacturer’s guarantee behind that, and one of the world’s strongest balance sheets waiting to absorb the asset because it is short of exactly these chips. In the base case, the rent retires the debt and none of the protections are tested, and in most versions of the tail, the layers absorb the loss as intended. The buildout looks precarious in isolation and considerably more durable once the incentives and the asset are examined together, which is the opposite of the conclusion the alarm draws from the same facts. The fragility that remains is narrow, living in the correlated downturn where the strong hands weaken together, and in the simpler deals that never spread their risk at all. That is what a well-built structure reveals: that the danger in this cycle is concentrated where the financial engineering is absent rather than where it is most advanced.

Kevin Warsh wants to elevate the role of the bond market in the determination of monetary policy.

The Federal Reserve has spent two decades training investors to expect that rate decisions will be telegraphed weeks in advance and that any change in direction will arrive only after a procession of carefully staged speeches. Markets no longer walk into an FOMC meeting asking what the central bank will do. They arrive knowing the answer and spend the afternoon parsing whether the statement’s language deviates slightly from the script. Warsh believes this predictability has corroded price discovery. Knowing the outcome of every meeting in advance is not a sign of a healthy market. It is proof that central-bank choreography has displaced the repricing function that Treasury markets are supposed to perform.

The Illusion of Control

Forward guidance began as an emergency response to the zero bound. With official rates pinned near zero, the Fed promised that rates would stay low for an extended period, hoping to drag down longer-term borrowing costs and encourage households and businesses to spend. The emergency tool hardened into a permanent operating principle, and the central bank moved from setting the overnight rate to attempting to manage the entire expected path of interest rates. The Fed, like all forecasters, does not know where inflation, unemployment, or growth will be twelve months from now, and it has demonstrated this repeatedly. Modeling the future trajectory of the economy has historically been a thankless task, especially as the economy evolves under the weight of decades of relentless innovation. Many of the models used by Fed economists were developed in the 1950s and 1960s when the US economy was less global, analog, and manufacturing-led, a much more cyclical environment than the service-driven framework we find ourselves in today. Forecasting is difficult enough in stable economic conditions.

Federal Reserve forward guidance converts those uncertain projections into something markets treat as a commitment, and Treasury yields, mortgage rates, corporate borrowing costs, and equity valuations are then constructed around a policy path that was never as reliable as investors were encouraged to believe. When the forecast changes, the Fed changes with it, and the market discovers that yesterday’s guidance was a conditional opinion dressed up as institutional certainty.

The Fed Controls Less Than It Admits

The deeper problem is that the Fed’s influence over inflation and employment is weaker than the institution concedes. Interest rates can affect credit creation, housing turnover, asset prices, and the appetite of leveraged companies to borrow. They cannot build power stations, produce natural gas, repair broken supply chains, remove tariffs, or expand the pool of skilled workers. Monetary policy cannot reverse the impact of automation on white-collar employment, and it cannot force hyperscalers to slow an infrastructure program they regard as essential to their survival. The forces increasingly driving inflation and employment are supply-side: energy scarcity, fiscal transfers, trade restrictions, immigration, and technology sit beyond the reach of the federal funds rate. The Fed can weaken demand elsewhere in the economy in response to a supply shock, but that is a blunt instrument that punishes the rate-sensitive sectors while leaving the source of inflation untouched.

This is the argument I made in Growth Without You. Economic growth is becoming less dependent on labor, credit expansion, and monetary stimulus, with productivity replacing employment as the marginal driver of output. AI can lift GDP, margins, and corporate profitability while simultaneously reducing demand for human cognition, which means unemployment can rise without a recession while inflation falls because the cost of producing goods and services declines. The Fed’s traditional playbook has no obvious answer to that combination, and guidance from an institution losing its grip on the variables it forecasts is worth even less.

Volatility Is Inevitable

Remove the guidance and volatility rises, which is unavoidable and, up to a point, desirable. The two-year Treasury yield will become more responsive to inflation, employment, and activity data. The distribution of possible outcomes around each FOMC meeting will widen, options markets will demand more compensation for uncertainty, and longer-dated bonds may carry a higher term premium because investors will no longer assume the Fed intends to hold their hand through every turn in the cycle. None of this means yields must rise continuously, and weak data could produce larger rallies because the market is no longer anchored to a pre-announced path. The defensible conclusion is that moves in both directions become larger. Uncertainty was always present; forward guidance merely suppressed its market expression, and Warsh appears comfortable watching it return.

The reality is the bond markets have not been volatile since QE became the prevailing policy tool in the previous decade. The average range for the US 10-year bond, peak to trough, since 2000 has been 133bps. Since 2012, only four years have produced a range in excess of the mean, with eight years experiencing ranges of less than 100bps. Lower absolute yields are a meaningful cause of the lower yield ranges, and bond price volatility would tell a slightly different story, but the message is clear. Long-duration Treasury yield swings have been muted when compared to prior cycles, and this is reflected in lower annual ranges, realized volatility, and poor Sharpe ratios for bond trading. While there is never just one answer, forward guidance definitely played a role.

Higher Yields Carry Less Information Than Warsh Believes

His confidence in the signal contained in higher long-term yields is harder to defend. The textbook assumption is that rising yields tighten financial conditions, weaken housing, restrain consumption, and eventually lower inflation. Some truth survives in that sequence, but the transmission mechanism has decayed. Fixed-rate mortgages are the clearest example. Higher Treasury yields hit new buyers, refinancers, and households that need to move, while leaving untouched the monthly payments of homeowners who locked in cheap thirty-year loans. Rather than producing an orderly housing slowdown, higher rates have frozen turnover, with existing owners refusing to surrender low-cost mortgages while prospective buyers face elevated financing costs and thin supply. The result is reduced activity without the price reset policymakers expect.

The corporate sector tells a similar story. The largest incremental issuers in the investment-grade market are increasingly the hyperscalers and the companies at the center of the AI buildout, and while these firms are not indifferent to the cost of capital, they are far less sensitive to it than the conventional model assumes. Microsoft, Alphabet, Amazon, and Meta hold enormous cash balances and strategic reasons to secure compute, data centers, electricity, and network capacity before demand is fully visible. A higher coupon will not persuade them to abandon projects they believe will determine competitive leadership for the next two decades. Borrowing capacity across these firms remains deep, and the willingness to invest is governed by strategic necessity rather than by modest moves in long-term yields.

The Press Conference Problem

The press conference post each FOMC meeting has become the contradiction sitting at the center of this policy transition. Warsh cannot appear before the press and refuse to discuss the future, because every serious question is forward-looking. Reporters will ask whether inflation is improving, whether employment is weakening, and what would cause the Fed to move at the next meeting. Even an answer framed around risks and incoming data moves markets, because investors immediately translate it into probabilities for future policy. The distinction between explanation and guidance is therefore less clean than it first appears: a chair cannot explain the current decision without revealing something about how the institution views the months ahead, and the tone of the answer, the risks emphasized, and the conditions attached to future action all become guidance whether he intends them to or not.

Chairman Warsh therefore has three realistic choices. He can keep holding press conferences and accept that markets will mine every answer for a policy signal, restrict them to meetings where rates change, or new projections are released, or eliminate them entirely and rely on written statements, speeches, and congressional testimony. After his last performance, fewer appearances look more likely, and that is the plausible reform. Limiting press conferences to meetings with updated projections or an actual policy change would reduce the volume of central-bank commentary without pretending that the Fed can discuss the economy publicly and avoid influencing expectations. The problem cannot be resolved through better wording. As long as the Federal Reserve speaks, markets will treat its words as guidance, and Warsh can reduce the frequency and precision of those signals but cannot eliminate them.

Be Careful What You Wish For

This leads to the question that should define his chairmanship: can the Federal Reserve tolerate the consequences of the volatility it creates? Silicon Valley Bank supplied an answer. In March 2023, higher yields exposed losses on long-duration assets and helped trigger a confidence crisis across regional banks, and once contagion became a risk, the Fed dropped $300bn of liquidity to stem the threat. The intervention was sensible, but it exposed the contradiction. The Fed tightened policy, generated losses across the banking system, and then supplied liquidity when those losses threatened financial stability. The institution that insisted markets absorb interest-rate risk stepped in the moment that risk threatened to spread, and Warsh inherits that reflex whether he wants it or not. He may prefer a cleaner market, but the White House, Treasury, Congress, and the rest of the FOMC will be less enthusiastic when cleaner pricing produces bankruptcies, falling asset values, or stress in strategically important sectors. Political tolerance for financial volatility is consistently overstated. Higher mortgage rates are acceptable until housing activity collapses, wider credit spreads are tolerable until companies cancel projects, bank losses are manageable until depositors question the safety of regional institutions, and equity declines are welcomed as healthy discipline until they threaten consumption, confidence, or an election.

Warsh’s timing to implement this strategy is awful. It has been five-plus years since US CPI has been at the 2% target. 30 Year Long Bond yields are at their highest level since 2007, and the combination of record deficits and hyperscaler issuance means that the supply of “high-quality” debt securities is vast. If the market, freshly appointed as judge, jury, and executioner, comes to the setting of rate policy and decides that the prevailing scenario is no longer sustainable, the United States could face its first-ever sustained period of rising risk premia. Rising yields and a steepening curve, driven by the idea that the issuance schedule cannot be absorbed, would be devastating for risk assets and risk-free assets alike, and a US equivalent of the UK’s “Liz Truss” moment cannot be ruled out. It is extremely unlikely, but when central banks leave it to the markets to determine where interest rates land, you introduce a level of risk that forward guidance did stem. The notion that free markets, if left to their own devices, can efficiently determine the price of money is nice in theory, but markets are not perfect, and oversight will be required to make sure that predatory behavior by well-capitalized market participants doesn’t manipulate interest markets for gain. There is an apparent hypocrisy between the actions of Kevin Warsh and coordinated Japanese Yen intervention. Allow markets to run their course to a point should be the mantra of US economic stewards. Kevin Warsh may be no different.

Honesty Is Not Effectiveness

Whether any of this makes policy more effective is a different question, and there is little reason to believe it will. The Fed’s difficulty does not stem primarily from poor communication. It is a weakening relationship between interest rates and the economy they are supposed to steer. Fixed-rate mortgages shield existing homeowners, cash-rich hyperscalers invest straight through modest changes in borrowing costs, supply-driven inflation sits beyond the reach of the funds rate, and AI-driven employment losses will not be reversed by cheaper money. Greater uncertainty may produce a more honest Treasury market, but honesty and effectiveness are not the same thing. Doubt about the path of rates can be shifted or concealed, never eliminated, and stripping it out of Fed communication will simply expose risks that were always present, with some borrowers struggling while the companies driving the AI investment cycle keep spending because they can afford to. The reform will be judged not by the first rise in yields but by what happens when higher volatility produces a casualty. At that point, Warsh will face the same choice as every modern Fed chair: allow markets to impose discipline and accept the economic consequences, or intervene and protect the system from the very risk he wanted investors to price. The Fed can create volatility whenever it chooses. We are about to discover whether it can tolerate the consequences.

Subscribe now

When I was writing Growth Without You, the advice I received was consistent: serious topics demand a solemn tone. Every time I slipped in a joke or an overly personal anecdote, I risked the book not being taken seriously. So I played it straight. I am incredibly proud of that book, but anyone who knows me well knows that the book’s sober tone is not a true reflection of who I am. Which is why How to Win a Trade War landed on me the way it did. Soumaya Keynes and Chad Bown have written a book about a deadly serious subject that is engaging, whimsical, and at times downright funny, and it is more authoritative for it, not less.

I met both authors back when I was running the US-China Series, pre-COVID. I certainly don’t claim to know them well, but reading this book, you feel like you do. Their personalities are all over the page. Keynes is the Financial Times economics columnist and podcaster with the killer one-liners. Bown has held a number of government roles, from the White House to the World Bank to the State Department, and is now a senior fellow at the Peterson Institute; he has probably forgotten more about tariff schedules than most of us will ever learn. They have great chemistry and genuinely differing viewpoints, which is precisely what has made their work, individually and as a collaborative unit, essential reading for anyone assessing global trade, the US-China relationship, and the volatility of Trump administration trade policy. To see the dynamic in action, watch their interview with Jon Stewart on The Daily Show. At times, it’s tough to work out who is the iconic comedian.

A quick summary of the key takeaways:

The only winning move is not to play. Borrowing from WarGames, the authors argue the best trade war strategy is avoiding one altogether. If you must fight, “winning” simply means losing less than your rival: minimizing your own wounds rather than landing a knockout blow.

Tariffs feel better than they work. Tariffs are seductive because they look decisive, but they obscure the underlying problem while imposing real costs at home. The authors’ first message to Washington is blunt: study your enemy before you swing.

Trade deficits are structural, not moral. The US-China imbalance is rooted in Chinese underconsumption and American overconsumption, with Beijing’s distortions propping up firms that should fail and Washington unable to kick its borrowing habit. Tariffs alone fix neither.

The old order isn’t coming back. The stable, rules-based trading system of the late twentieth century is gone, and trust has eroded too far to simply restore it. Their prescriptions are refreshingly modest: safeguards, annual expectation-resets between the great powers rather than grand permanent settlements, and, most ambitiously, an international competition authority.

Heavy stuff and yet, the book never reads that way. It is littered with metaphors that constantly land: global imbalances as an unbalanced see-saw, with a heavyweight on one end and a lightweight on the other; trade deficits explained through video game consoles; the world’s great trading powers rendered as a fleet of mismatched ships. I won’t spoil the punchlines. The metaphors are ridiculous and they work, because underneath the whimsy is rigor. This is a quick read, but you spend enough time in the weeds to achieve the level of detail required for an accurate assessment of the fragilities of the global trade framework, and of what needs to be done, or frankly can’t be done, to fix it.

What makes the book sing, though, is the working relationship behind it. This is not a book where two voices get sanded down into one beige institutional tone. You can feel the collaboration on the page: the negotiations, the disagreements, the moments where one author clearly talked the other out of, or into, something. The small human details of the partnership come through too, right down to the almonds deployed to keep the “hangries” at bay during the writing process. As someone whose wife eats almonds for precisely the same reason, I can appreciate this fully. Nothing good gets written, or negotiated, on an empty stomach, a lesson that applies equally to co-authors and to superpowers. It is exactly why the book works: the anxious global economy of their subtitle is navigated by actual humans, and they never let you forget it.

I listened to the audiobook while reading along. Definitely worth double dipping into both.

I would still love you all to read Growth Without You on the beach this summer. But I have to concede: How to Win a Trade War is going to end up covered in more sand than my serious work, and having read it, I understand why.

In 2013, the world learned about Chinese “Ghost Cities”. 60 Minutes filmed the empty towers of Zhengzhou’s new districts and told America that China had built cities nobody would live in. It was a sign of property excess and the failure of centralized planning. It caused an investor panic, and I got completely wrapped up in it. I made a trip to Shanghai, spoke with mainland contacts, and did everything in my power to position myself and my clients for the impending blowup. Within two years, the district where those towers stood reported over a million residents. I could have saved myself a lot of embarrassment had I had a conversation with a local Hong Kong property developer who explained China’s multi-trillion-dollar urbanization push simply.

You have to build an apartment before people move in.

That is where artificial intelligence stands in 2026. Some of the buildings are up, and the tenants are still moving in. The people pointing cameras at the empty floors are certain they are watching a bubble. The lessons of the evolution of the Internet and how companies utilized this foundational technology imply that the optimists are early rather than wrong.

The current bear case against corporate AI adoption rests on sound data, a snapshot of the current state of play, and a reputable MIT study. 95% of enterprise pilots produce no measurable P&L impact. The supporting evidence is deep, from earnings-attribution surveys to rising project abandonment, and even the most bullish amongst us have to take this data seriously.

The assumption underneath every one of those numbers is that real returns would be visible in company results by now, three years into deployment. In my book, Growth Without You, I argued that AI is following the same J-curve the internet traced, where returns arrive only after companies rebuild their processes around the technology, enhancing workers first and replacing them later. This report tests that claim against the record of 1995 to 2007, focusing on a bottom-up assessment of deployment. The internet, like AI, was a company-by-company success story rather than some magical macro event. The people demanding aggregate proof in 1999 were right about the quarter and wrong about the decade, and the absence of measurable ROI today is a timestamp, not a verdict.

The Bears Are Reading the Data Correctly

AI skeptics are accurately describing today’s environment. McKinsey’s most recent global survey found that 88% of companies now use AI somewhere in the business, yet only 39% report any earnings impact at all, and most of those put the contribution below 5% of EBIT. Deloitte’s survey of more than 3,200 executives tells the same story from a different angle: two-thirds report productivity and efficiency benefits, while only one in five can point to AI-driven revenue growth. The Census Bureau, running the largest American adoption survey, counts just 17% to 20% of American firms using AI to produce goods and services at all. S&P Global’s enterprise survey reported that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier.

However, it is the MIT study that underwrites broader doubts on AI’s efficacy. Released in August 2025 under the title The GenAI Divide, the research combined 150 executive interviews, a survey of 350 employees, and an analysis of 300 public AI deployments, and it concluded that only about 5% of enterprise pilots achieve rapid revenue acceleration while the rest deliver little to no measurable P&L impact. The headline number traveled through the world’s investment committees within weeks, usually stripped of two things: the caveats and the diagnosis. The caveats matter, because the report published no supporting data for its 95% and drew a public demand for retraction from Wharton’s Kevin Werbach.

The diagnosis is more valuable. The researchers attribute the failures to a learning gap rather than to the models, observing that generic tools stall inside enterprises because they do not adapt to existing workflows, that purchased solutions succeed at roughly twice the rate of internal builds, and that more than half of AI budgets flow to sales and marketing while the strongest returns appear in back-office automation. Read closely, the bears’ favorite study is an argument that ROI follows workflow redesign and eludes everyone else. Anyone who ran a corporate web pilot in 1999 would recognize every page.

Even after heavy discounting of the MIT headline, the survey record is consistent: enterprise AI, three years in, has not produced broad, measurable, attributable returns. The problem with this evidence is not its accuracy but its precedent. A poll of IT executives in March 2002 found that 68% rarely or never measured the return on their technology projects, and nearly three-quarters had no formal process for doing so. That poll was taken two years after the spending peak, at the bottom of the bust, seven years into the commercial internet. The 2026 numbers rhyme: KPMG finds barely half of enterprises have dashboards monitoring their AI costs, a third concede they scarcely understand token-based pricing, and BCG reports 94% of executives will keep investing at current or higher levels even if nothing pays off within a year. McKinsey puts the share of organizations tracking well-defined KPIs for their generative AI work below 20% despite calling KPI tracking the practice with the greatest bottom-line impact. Most companies did not measure technology returns then, and most are not tracking them now, yet the spending continues, with four out of five CEOs telling BCG they are more optimistic about AI returns than a year ago. The people running the businesses can see something the surveys cannot.

If a researcher had run the MIT study’s methodology across corporate America in 2002, the failure rate would have looked at least as damning, because most corporate web projects of the era were pilots that quietly died or portals nobody used. The companies that eventually built enormous returns from the internet could not have shown them to a surveyor in 2002. The returns were being assembled through trial and error within specific business processes, and the process took years.

Subscribe now

How Internet ROI Was Built

Selling airline tickets is the example I keep returning to. When Alaska Airlines sold the first US airline ticket over the internet in December 1995, distribution was one of the industry’s highest costs, with agent commissions ranking as the fourth-largest expense line after labor, fuel, and aircraft. Four years later, the web was still a rounding error at 1% to 2% of ticket revenue, and anyone investigating airlines about internet ROI in 1999 would have concluded the technology had failed. The unit economics said something else entirely. A booking through a travel agent cost an airline about $46 against under $24 through its own website, so the airlines rebuilt the channel around that arithmetic, cutting commissions year after year until Delta, which had spent $540 million on them in 2001, stopped paying altogether in 2002. By then online bookings had gone from 7% to 30% of tickets and the cost per booking had fallen by nearly half. Paper tickets, $10 to process against $1 for an e-ticket, were abolished worldwide in 2008. It took seven years to kill the old channel and twelve to finish the job, and the returns sat in the unit costs the whole time, years before any corporate assessment could see them.

Banking ran the same experiment on an even steeper curve. The Federal Reserve Bank of Chicago published the arithmetic in 2001: a branch transaction cost about a dollar, an ATM transaction cost about a quarter, and an internet transaction cost roughly a penny. A hundred-to-one cost advantage still took a decade to matter, because at the end of 1999 only a fifth of national banks offered internet banking and just fourteen million Americans had ever tried it. By late 2004, a quarter of American adults banked online. The payoff surfaced in customer economics rather than in anyone’s survey, with Bank of America finding that online bill-pay customers went from slightly less profitable than their offline peers to 21% more profitable within three years. Nobody who asked banks about internet ROI in 1999 found any. It was being assembled in the cost per transaction, one process at a time.

The earliest movers put the proof in their securities filings. Dell’s fiscal 2000 10-K shows nearly half of revenue coming through dell.com at $40 million a day, inventory at six days of supply, and a cash conversion cycle of negative 18 days, which means customers were financing the company’s growth. Cisco took 90% of its orders over the web by 2000. Schwab repriced the stock trade from roughly $80 to a flat $29.95 in January 1998, absorbed a single down quarter, and finished the year with revenue up 19% as online channels handled 80% of trades by 2000. GE said web procurement cut its transaction costs from $50 to $5. None of these returns shows up in an economy-wide survey of internet ROI, because each is a margin line inside a specific redesigned process, and that is precisely where the bears refuse to look today.

Rebuild the Factory Floor

The pattern of technology adoption has a mechanism, and it was documented before the internet existed. Robert Solow wrote in 1987 that you can see the computer age everywhere but in the productivity statistics, and the economic historian Paul David answered him with electricity. Electric motors powered less than 5% of American factories in 1900, two decades after Edison’s lightbulb, because manufacturers installed dynamos as drop-in replacements for steam engines and gained almost nothing. The payoff came only when engineers redesigned the factory floor around distributed power, at which point total factor productivity in manufacturing jumped from 0.5% a year to 5.3% a year in the 1920s. Erik Brynjolfsson measured the same lag for computing: across 527 large US firms, the returns visible over five-to-seven-year horizons were up to five times larger than what showed in the first year, because every dollar of hardware required roughly nine dollars of organizational investment that never appears on a balance sheet. The current absence of ROI that the Bears catalog is not evidence against the technology. It is a long implementation window that is observed in every deployment of a technology of this kind.

The top-down confirmation arrived, but it came last. US labor productivity growth ran at 1.4% per year from 1974 to 1995, accelerated to roughly 2.5% in the late 1990s, and peaked near 3.5% between 2000 and 2003, after the Nasdaq crash. McKinsey’s much-quoted 2001 study concluded that just six of sixty sectors had produced 99% of the late-1990s acceleration, and for years, it was cited as proof that the internet’s benefits were narrow. The margin data buried that conclusion. The S&P 500’s net profit margin climbed from 3.4% at the end of 2001 to 8.6% by the end of 2006, corporate profits after tax rose from 4.8% of GDP to 10.8% over the same five years, and the Fed’s own industry decomposition found the post-2000 gains had moved beyond the technology producers. While the margin gains were not just attributed to the internet with globalization playing a leading role, we can draw some general conclusions. Adoption broad enough to double the profit share of the entire economy is not a six-sector story. The people who called the internet a failure in 2002 were describing the statistics accurately but completely misreading the breadth. Getting a mega theme wrong is usually not a failure of observation. It is a failure of patience with the mechanism.

AI Is Running the Identical Experiment

Skepticism about the business impact of new technology is as old as machinery itself. It starts at the loom, runs through the factory owners who saw no reason to scrap working steam engines for electric motors, and arrives at every desk that ever hosted a computer the accountants could not justify. Each generation considers its own doubt uniquely sophisticated, and each generation gets a version of the same essay. The internet’s version landed in May 2003, when the Harvard Business Review published “IT Doesn’t Matter,” declaring corporate technology spending a commodity cost with no competitive payoff and observing that the companies with the highest returns were the ones investing the least in it. The timing was exquisite, because the productivity statistics were peaking and the margin expansion of 2002 to 2006 was already underway as the article hit desks. Doubt of this kind is not stupidity; it is what a payoff lag looks like from the outside.

The AI record already holds both of the internet era’s outcomes, redesigns that work and experiments that fail, and the same mechanism separates them. Salesforce reports $100 million in annualized support cost savings, with AI agents handling 3 million support conversations. Marc Benioff is on record as saying that support headcount has fallen from 9,000 to roughly 5,000. Jamie Dimon says JPMorgan gets “about $2 billion of benefit” on $2 billion of AI expense, with 150,000 employees using the bank’s internal model weekly, which is exactly what year three of a GPT deployment should look like: breakeven at a well-run institution, not a bonanza.

On the other side of the ledger, McDonald’s ended its three-year drive-thru AI pilot with IBM, Air Canada’s chatbot invented a bereavement policy that a tribunal made the airline honor, and Commonwealth Bank of Australia cut 45 service roles on a chatbot’s promise, then admitted the error and reversed the cuts. For every success, there are glaring failures. Klarna compressed the whole dynamo cycle into fifteen months, announcing in February 2024 that its assistant did the work of 700 agents and projecting a $40 million profit improvement. Then, by May 2025, it conceded that cost-cutting had gone too far and rehired humans for quality. These are not contradictory data points. They are what trial and error looks like from inside.

The diagnostic evidence says the variable is redesign, not the technology. The MIT learning gap said as much. McKinsey finds that fundamentally redesigning workflows is the practice with the strongest link to AI’s impact on EBIT, and that its high performers are nearly three times as likely to have done so. BCG tells its clients that successful AI programs are 10% algorithms, 20% data and technology, and 70% people and process. That is the factory floor argument restated as consulting speak. The companies dropping AI into existing processes are collecting Air Canada and Klarna outcomes, while those rebuilding the process around it will eventually achieve Salesforce results, the same split the dynamo created in 1910 and the web repeated in 1999.

Blowing Through the Budget Is the Point

Somewhere inside every large company on earth, there is a room full of AI nerds having the time of their lives, burning through underpriced inference in search of workflows worth rebuilding. Uber exhausted its entire 2026 budget for AI coding tools in four months, having encouraged the binge with an internal leaderboard, and its chief operating officer admits that the link between token spend and consumer benefit “is not there yet.” Microsoft canceled most of its internal Claude Code licenses on cost grounds six months after issuing them, while Meta and Amazon kept their own usage leaderboards. Shopify, Spotify, ServiceNow, and Roku have all told earnings calls that AI is surging as a share of operating expenses. The consumption data behind the anecdotes points one direction, with OpenRouter routing 25 trillion tokens a week this spring, five times its volume of six months earlier, Google’s direct API token usage growing 60% in a single quarter, and Deloitte finding two-thirds of enterprises now consume over a billion tokens a month. Ramp’s payments data says the average business is spending thirteen times more on AI tokens than it did in January 2025.

None of this is what failure looks like. A company that blows through its annual inference budget in four months is not a company that considers the technology useless; it is a company whose engineers found so many things worth trying that finance could not keep up. This is the 2026 edition of the corporate web team of 1999, the group in the basement rebuilding the ordering system and the customer database while the CFO wondered where the money was going. The trial and error is the point, because nobody knows in advance which workflows the technology transforms, and the only way to find out is to spend on tokens. The airlines did not know in 1996 that distribution would be the redesign that mattered, and Uber cannot yet tell which of its experiments will, which is why its executives’ honesty about the missing link is evidence of process rather than proof of waste. Budget overruns at this stage of a general-purpose technology are the price of admission.

Margins First, Surveys Last

The margin expansion is no longer hypothetical, which is the part the bears have to explain away. Salesforce says its support costs are down 17% since early 2025, and ServiceNow told investors that AI reasoning is less than 10% of its cost to serve, that subscription gross margins are sustainable above 80%, and that agentic AI should deliver $300 million in annualized savings. The heaviest internal deployers are printing the pattern at scale, with Meta’s first-quarter net income up 61% and Alphabet’s up 81% against revenue growth in the thirties, which is what margin expansion looks like on an income statement. In my June report And the Answer Is Margins, I noted that the S&P 500’s blended net margin reached 14.7% in the first quarter of 2026, the highest on record, with seven of eleven sectors expanding. I had no proof then that AI was the proximate driver of a single quarterly print, and I still do not, but the direction aligns with the composition I laid out in Growth Without You, where 60% to 70% of early AI gains accrue as cost savings rather than new output. Deloitte’s surveys say the same. Margins are where the returns show up first, exactly as unit costs were where the internet surfaced.

The workforce arc explains the timing of everything else. Today’s tools enhance far more than they replace, which is why the Federal Reserve finds 41% of American workers already using generative AI while only 18% of firms have formally deployed it, and why Dimon talks about retraining while Benioff redeploys support staff into sales rather than out the door. My base case is that current models can displace only the bottom tier of cognitive work, perhaps the least productive tenth, because they are not yet reliable enough to run unsupervised, which is the lesson Klarna paid to learn in public. That ceiling will not hold. As the models improve and workflows are rebuilt around them, substitution broadens the skill distribution, headcount follows the redesign rather than precedes it, and margin expansion compounds from a support-desk story into an income-statement story. That is when the impact is truly felt, first in earnings and soon after in the labor market; the sequence Growth Without You maps in full. The bears are measuring today’s tools against tomorrow’s claim.

The indicators worth tracking are bottom-up ones: the share of companies attributing EBIT impact in the McKinsey series, disclosed unit-cost lines like Salesforce’s, workflow-redesign adoption, and whether reversal cases rerun successfully, as McDonald’s says a drive-thru voice solution remains in its plans. Brynjolfsson, the economist who measured the computing lag, wrote in February that with US productivity growth near 2.7% in 2025, the data suggest “we are now transitioning out of this investment phase into a harvest phase.” Surveys of averages will be the last place the returns appear, as they were in 2002.

The precedent carries one warning, and the airlines are it: they cut distribution costs per booking by 43.6% in three years, yet, in my reading of the two decades since, competition passed most of those savings to passengers rather than shareholders. The gains persist where structure lets a company keep them, which is Warren Buffett’s old point about GEICO, that “being the low-cost producer is all-important” and the advantage compounds only when rivals cannot instantly match it. The durable winners will be adopters whose cost advantages are hardest to copy, spread across thousands of ordinary companies, which is the case I made in And the Answer Is Margins for owning the equal-weight S&P 500 against the cap-weighted index. The 493 are today’s airlines and banks in 1999, holding redesign gains no survey yet measures, and the margin expansion belongs to the tenants rather than the landlords.

Conclusion – Repeating past bearish mistakes

The question was how long it took companies to build ROI from the internet, and the answer is specific: the unit-cost evidence appeared inside three to five years at the companies that redesigned processes around the technology, the P&L consequences compounded over five to ten, and the old channels were dead within a decade, while the surveys and statistics confirmed all of it only after the fact. AI three years in looks exactly like the internet three years in, including the failed pilots, the premature headcount cuts, the walked-back announcements, and the handful of companies already posting nine-figure savings because they rebuilt the workflow instead of renting the tool.

Predicting the failure of a technology during its trial-and-error phase is a tradition with a distinguished roster. Newsweek declared the Internet “baloney” in 1995, asking how online commerce could ever work when “my local mall does more business in an afternoon than the entire Internet handles in a month,” and Paul Krugman wrote in 1998 that “by 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.” By 2005, the S&P 500’s margins were making new highs on the redesigns those years of trial and error had paid for. MIT’s 95% and the abandonment surveys are the 2026 editions of the same complaint, accurate about the present and silent about the mechanism. The absence of measurable ROI in the averages is a snapshot of the prevailing stage in a long evolutionary cycle, not a verdict, and the companies assembling returns right now are doing it in unit costs where surveys do not look. The investors who understood this in 2002 bought the adopters and were paid for a decade. The ones who demanded aggregate proof first eventually got it, years later and at full price.

The market will not wait for ROIs to appear in quarterly filings before re-rating companies designing workflows around AI. The market will eventually reward those companies brave enough to experiment because every failure is one step closer to appreciable margin expansion.

Subscribe now

One of the most frustrating things about publishing a book is realizing, two weeks later, that you should have included one more idea. In Growth Without You, I spend a lot of time on the fiscal sustainability problem that artificial intelligence creates for the United States. The argument was not that Washington spends too much money, although it does, or that entitlement reform will be easy, because it will not be. The deeper problem is that the U.S. tax system was built around labor income at precisely the moment when AI is beginning to reduce the economy’s dependence on labor. That was the point I made in the book, and I still believe it is correct. What I should have added is that the tax code does not merely fail to adapt to artificial intelligence; it actively pushes companies toward replacing human cognition with software.

The United States has a tax base problem hiding in plain sight. In 2025, the U.S. economy generated roughly $13 trillion in wages and salaries, compared with about $4.1 trillion in corporate profits. Labor remains the larger pool of income, but the direction of travel is what should worry policymakers. Profits are growing faster than wages, corporate margins are expanding, and AI gives companies a way to increase output without adding workers at the same rate. The government taxes the slow-growing pool of income heavily and automatically. It taxes the faster-growing pool more lightly, with more room for timing, deductions, deferral, and avoidance.

The arithmetic already looks uncomfortable. The IRS collected roughly $1.95 trillion through individual income-tax withholding and another $1.58 trillion through FICA in fiscal 2025. FICA stands for the Federal Insurance Contributions Act, the payroll-tax system that funds Social Security and Medicare.

That means roughly $3.54 trillion was collected through the labor-payroll channel. Against a wage base of roughly $13 trillion, the effective federal tax burden on labor is near 27%. Federal corporate income-tax collections were about $484 billion, which is a fraction of what Washington collects from wages, salaries, and payroll-linked taxation.

The structure of FICA is central to the argument. In 2025, Social Security tax applied to wages up to $176,100. The employee paid 6.2%, and the employer paid another 6.2% on the same wage base. Medicare tax had no wage cap, with the employee paying 1.45% and the employer paying another 1.45% on all covered wages. For most workers below the Social Security wage base, the combined FICA burden was 15.3% of wages, split between the worker and the company. The employer-side cost alone added 7.65% to the wage bill before healthcare, training, compliance, office space, or management time entered the calculation.

The capital-gains and dividend question needs to be handled carefully. Tax Policy Center estimates that the federal government collected roughly $271 billion in individual income tax from long-term capital gains and qualified dividends combined in 2025. That number is meaningful, but it does not change the comparison because it remains far below the $3.54 trillion collected through the labor-payroll channel. Treasury’s separate 2025 estimates are also important, but they measure foregone revenue from preferential rates rather than taxes paid. Treasury estimates that preferential treatment for qualified dividends reduced federal revenue by about $40 billion in 2025, while preferential treatment for capital gains reduced revenue by about $119 billion. These numbers reinforce the point that capital income receives more favorable treatment, but they should not be treated as equivalent to corporate income-tax receipts.

Capital-gains taxes should not be included in the main comparison because they are not directly linked to annual corporate profits. They are taxes on realized asset appreciation, shaped by market prices, holding periods, and investor timing decisions. Qualified dividends are closer to the profit stream, but even there the tax is paid by shareholders rather than companies and depends on payout policy and investor ownership structure. This is not a perfect comparison, and it is open to debate, but I think the cleanest comparison is between the labor-payroll channel and federal corporate income-tax receipts. The former is automatic and tied to employment. The latter is direct taxation of corporate profit, which is the income stream AI is likely to expand.

That comparison still understates the distortion because it treats the issue only as a revenue problem. The more important point is the incentive embedded in the tax code itself. A company that hires a person pays wages, benefits, employer payroll taxes, unemployment insurance, workers’ compensation, compliance costs, and administrative overhead. A company that deploys AI pays a vendor, buys software, rents compute, or builds internal systems. The task may be economically similar, especially as AI improves, but the tax treatment is not. Human cognition arrives with payroll taxation attached, while artificial cognition does not trigger FICA at the point of deployment.

Assume a company has a role that costs $100,000 in wages. Before healthcare, training, office space, management time, and compliance, the employer owes another $7,650 in payroll taxes. The employee also has $7,650 withheld, reducing take-home pay and funding Social Security and Medicare. Now assume the same company can buy an AI system for $100,000 that performs enough of the role to make the human position unnecessary. The company pays the invoice and avoids the employer-side payroll tax attached to the worker. The government may collect tax somewhere else in the chain, but the payroll tax connected to that job has disappeared.

And that is before any allowance for the assumed cost savings of AI versus the current cost of labor.

That is not a neutral tax system. It is a tax system that penalizes human employment and leaves the software substitute outside the payroll-tax base. Companies already have operational reasons to adopt AI: speed, consistency, scalability, and lower marginal cost. The tax code adds another reason by making people more expensive than software before any productivity comparison is made. This does not mean every job is at immediate risk, and it does not mean artificial cognition is equivalent to human cognition today. It means that as AI improves, the tax code will increasingly tilt substitution decisions against workers.

The unfairness sits beneath the productivity story. A human worker competing with artificial cognition is not competing only on ability, output, reliability, or cost. The worker is competing against a tax system that attaches a funding obligation to employment but not to software. If two inputs can perform the same task, and one carries payroll taxation while the other does not, the playing field is not level. The tax code is not passively observing the transition from labor to capital. It is helping finance the transition by making labor the more heavily taxed input.

The defenders of the current system will argue that AI is not tax-free. They are right, but that does not solve the problem. The software vendor may pay corporate income tax, its employees may pay payroll taxes, its shareholders may eventually pay taxes on dividends or realized capital gains, and the data center may pay property or electricity-related taxes. None of that changes the decision facing the company replacing a person with software. At the point of substitution, the hiring firm removes a worker from payroll and removes the FICA burden attached to that role. The tax base does not vanish entirely, but it moves into channels that are less automatic, less labor-linked, and usually more favorable to capital.

This distinction is essential because fiscal policy is still built around the paycheck. W-2 income is visible, recurring, and collected every two weeks before the worker ever sees the money. Capital income is mobile, easier to defer, and more exposed to legal structuring. Payroll taxes fund Social Security and the hospital insurance portion of Medicare, which means the system depends on employment remaining central to the economy. AI challenges that foundation directly. If output rises while payroll employment grows slowly, the government can have GDP growth, profit growth, and weakening social-insurance revenue at the same time.

That is the fiscal trap AI creates. The economy can look healthy in aggregate while the tax base deteriorates underneath it. GDP can rise, margins can expand, and equity markets can grind higher, even as the government’s most reliable revenue channel grows more slowly. This is not the pattern of a conventional recession, where tax receipts fall temporarily and recover when employment returns. AI displacement is different because the task may not return when demand recovers. If software performs the role faster, cheaper, and with no payroll-tax obligation, the old job becomes a cost structure that management has little reason to rebuild.

The result is a feedback loop that policymakers are not prepared to confront. Companies adopt AI to improve productivity and protect margins. The adoption reduces labor intensity and weakens payroll-linked revenue. The same substitution increases profits, but the tax code captures those profits less efficiently than wages. Deficits widen, not because the economy is weak, but because the income mix has changed. Washington then faces rising transfer pressure from displaced workers while the most reliable tax channel becomes less able to fund the promises already made.

This is why the Federal Reserve is almost irrelevant to the problem. Rate cuts can support asset prices, ease financial conditions, and prevent liquidity stress from becoming systemic. They cannot make a company rehire a worker whose function has been absorbed into software. They cannot rebuild the payroll-tax base if output is increasingly generated by capital rather than labor. The Fed was designed to manage cyclical weakness in a labor-driven economy. It has no instrument designed to correct a tax code that makes human work more expensive than artificial cognition.

The policy answer should begin with neutrality between human and artificial cognition. If the tax code wants to fund Social Security, Medicare, and the federal state through the productive economy, it cannot place most of the burden on employment while exempting the software that replaces employment from payroll taxation. That does not require a cartoonish robot tax, and it does not require punishing every company that adopts AI. It requires admitting that payroll taxation was designed for a world where labor was the main productive input. Once capital and software perform a larger share of the work, the tax base has to follow the work.

The cleanest reform is to reduce the tax penalty on labor while increasing the burden on capital-linked income. Ordinary income, qualified dividends, and corporate profits cannot be treated as separate worlds when AI is shifting income from workers to owners. Capital gains belong in a different category for the purpose of this calculation because realization is discretionary and not directly tied to current-year profits. That does not mean capital gains should escape reform, and it certainly does not mean the preferential treatment of realized gains is defensible in a capital-heavy economy. It means the central comparison in this report should remain focused on the automatic taxation of labor versus the direct taxation of corporate profits. That is where the employment incentive is most visible.

Personal income-tax rates could come down if the effective tax rate on corporate profits and distributed capital income rose meaningfully. Corporate taxation should be judged by what companies actually pay, not by headline statutory rates that bear little resemblance to the final transfer to the state. Dividend taxation should be part of the broader debate because distributed profits are one of the ways corporate earnings reach owners. Payroll taxation should be redesigned so the burden of funding social insurance does not fall overwhelmingly on human work. The objective is not anti-business. The objective is to stop taxing the input being displaced while under-taxing the input doing the displacement.

Corporate America will resist this because the current system works beautifully for capital. Companies get the productivity upside from AI, shareholders receive the margin benefit, and the government absorbs the revenue loss from a weaker payroll base. That bargain cannot survive indefinitely. A country can tolerate high profits, rising asset values, and aggressive automation if the gains are broadly taxed and politically recycled. It cannot tolerate a system where companies are rewarded for replacing workers, workers fund the state until they are displaced, and the software substitute carries none of the same social obligations. The current tax code may have been defensible in a labor economy, but it becomes indefensible in an AI economy.

The political consequences will not arrive as a clean policy debate. They will arrive as anger, populism, resentment, and hostility toward the firms that benefit most from automation. Investors should not dismiss that risk as moralizing from people who do not understand productivity. Capital’s current dominance depends on a stable society, a functioning state, and a tax system viewed as legitimate enough to survive democratic pressure. If AI allows companies to produce more with fewer people while the fiscal burden remains on labor, the legitimacy of the system will weaken. The danger is not that capitalism becomes too profitable, but that it becomes too visibly detached from the workers and taxpayers expected to support it.

For decades, I believed America’s fiscal problem was manageable because the country’s growth engine was stronger than Washington’s dysfunction. I still believe in the strength of U.S. corporate America, and I remain convinced that AI will reinforce many of America’s economic advantages. What has changed is my view of the tax base beneath that advantage. A labor-based tax system cannot fund a capital-heavy economy, and it certainly cannot survive if the tax code rewards the replacement of taxable workers with untaxed software. The United States does not need to stop AI adoption, which would be impossible and economically self-defeating. It needs to stop pretending that taxing human cognition while exempting artificial cognition is a neutral way to fund the state.

Subscribe now

Edmund Morris’s three-part epic on the life of Theodore Roosevelt has few biographical peers. While all must tip their hat to Robert Caro’s exploration of the life of Lyndon Johnson, Morris’s account of the extraordinary life of the first great presidential progressive is made all the easier by the richness of a life of only sixty years. Be it as a big-game hunter, explorer, conservationist, historian, war hero, or author of three dozen books, his life outside the presidency was arguably more interesting than his two terms after the assassination of William McKinley and his failed 1912 run as a third-party candidate. He survived his own assassin’s bullet that year in Milwaukee while campaigning, shot in the chest, and only saved by a 50-page speech and his steel eyeglass case. Before going to the hospital, he delivered that hour-long speech to adoring crowds, opening with the immortal line.

“Ladies and Gentlemen, I don’t know whether you fully understand that I have just been shot, but it takes more than that to kill a Bull Moose.”

The Republican Party of Theodore Roosevelt is unrecognizable compared to what we see today. The Democratic Party of that era is just as difficult to map onto the present, given its dependence on Southern segregationists and regional power brokers whose politics bear little resemblance to the modern party. Both parties have changed dramatically over the past 150 years, and attempts to draw straight lines of policy continuity across that period quickly collapse. The labels may be the same, but the economic priorities, regional coalitions, and governing assumptions have been remade several times over.

Roosevelt belonged to a Republican tradition that accepted an active federal government as a necessary counterweight to concentrated private power. He saw national markets, corporations, and supply chains as requiring Federal oversight. This was not hostility to capitalism. It was a belief that capitalism needed rules, supervision, and public accountability if it was going to retain legitimacy.

That worldview sits uneasily with the Republican Party of Donald Trump, where suspicion of the administrative state is central to the political project. Roosevelt’s Republicanism moved in the opposite direction. He viewed monopoly as a democratic problem, federal supervision as a necessary tool, and property rights as subordinate to human welfare when the two came into conflict. There would be no obvious place for Colonel Roosevelt in a modern Republican administration. His politics were built around disciplining concentrated power on behalf of the broader republic. That instinct, more than any single policy difference, separates his Republicanism from the party that carries the label today.

The central economic challenge of the early 20^th century was not the size of government, the deficit, or the culture war. It was the trusts.

By the early twentieth century, Standard Oil, U.S. Steel, American Tobacco, and the major railroad combinations had evolved beyond the boundaries of ordinary corporate success. They were not simply large companies operating in competitive markets. They had become business empires that controlled the infrastructure of American capitalism itself, setting the terms on which oil was refined, steel was produced, freight was moved, and consumers accessed essential goods. Their reach extended across state lines, while the political tools available to restrain them remained local, fragmented, and inadequate. This was the core imbalance of the Progressive Era: private enterprise had scaled nationally, while democratic oversight had not. Roosevelt and his presidential successor and then rival, William Taft, were forced to confront a reality that every generation eventually faces in its own form. Capitalism had produced private organizations powerful enough to overwhelm the public framework meant to govern them.

Roosevelt’s answer was not a blanket assault on business. He did not believe that every large corporation was dangerous simply because it was large. He understood that scale was part of modern industrial capitalism. Taft, despite being remembered less vividly than Roosevelt, pursued antitrust enforcement with equal seriousness and, in some respects, greater legal effect. The case against Standard Oil, initiated under Roosevelt and concluded under Taft in 1911, became the defining example of the federal government asserting that even the most powerful private enterprise could not stand above the public interest.

In Roosevelt’s time, the issue was oil, steel, tobacco, and railroads. In ours, it is artificial intelligence. The technologies are different, and the economy they operate in is unrecognizable. But the underlying question is familiar enough to deserve revisiting.

How would Teddy Roosevelt have dealt with AI?

AI is the New Trust Problem

The comparison between Roosevelt’s attempts to rein in the trust and artificial intelligence should not be treated as a perfect historical parallel. Standard Oil, U.S. Steel, American Tobacco, and the railroad combinations were products of an industrial economy built around physical assets. They controlled refineries, pipelines, freight routes, production capacity, and distribution networks. Their power came from the ability to dominate the channels through which commerce moved.

AI represents a different form of concentration, but the economic significance is just as profound. The better comparison is not between railroads and cloud computing, or between oil refineries and data centers. It is between two eras in which private power began to challenge the foundations that allow capitalist societies to function.

In Roosevelt’s time, the trusts threatened competition. They could crush rivals, dictate terms to suppliers, restrict market access, and use scale to convert private advantage into structural dominance. AI presents a different but related challenge. It redefines the role of the worker by reducing the need for human judgment across large parts of the economy and shifting bargaining power further toward capital.

Competition and labor are not peripheral features of capitalism. They are central to its legitimacy. Competition enables private enterprise to remain dynamic rather than entrenched. Labor ensures that growth is connected to wages, mobility, and broad-based participation in the economy. The trusts weakened the competitive pillar. AI threatens labor’s link to broader prosperity.

This is why the Roosevelt framework is relevant again. He did not believe large corporations were inherently dangerous. He accepted that scale could produce efficiency, lower costs, and strengthen national power. His concern was the point at which scale became coercive, when private companies used their position to suppress competition, distort politics, exploit consumers, or push the costs of their dominance onto society.

The same standard should be used to judge AI. Good AI raises productivity, improves medicine, strengthens education, reduces waste, accelerates scientific discovery, and expands national capacity. Bad AI entrenches monopolies, captures data, manipulates information flows, weakens career ladders, displaces workers, and leaves taxpayers to absorb the social cost while shareholders capture the gains.

Roosevelt would not have viewed corporate scale as the problem. The danger emerged when private concentration stopped serving the public and began exploiting it. Applied to AI, the concern is not the inevitable growth of the leading platforms, but whether that growth remains anchored to the public good or becomes another private tollbooth sitting across the economy.

Work as Central Economic Pillar

The central issue for Roosevelt would be unemployment. He did not view work as a narrow economic input. Work was tied to dignity, family security, self-respect, and social stability. It was how people supported households, raised children, built communities, and felt connected to the national project. A capitalist economy can absorb disruption. A democratic society struggles when millions of people conclude that the system has no meaningful place for them.

This is where AI becomes a more profound challenge than the trusts. The old trusts threatened competition. AI threatens the role of labor. The promise for companies is obvious: higher output, stronger margins, faster growth, and less dependence on hiring. For investors, this is an extraordinary proposition. For society, it raises far more difficult issues.

The uncomfortable truth is that productivity gains no longer guarantee labor gains. That assumption sits behind most of modern economic policy. Companies become more efficient, profits rise, investment follows, and workers eventually benefit through higher wages, better jobs, or new opportunities. AI weakens that chain. The gains can accrue first to margins, market capitalization, and capital owners, while the labor market adjusts slowly and painfully in the background.

This is the Growth Without You problem. The economy can expand, profits can rise, and equity markets can rally even as the demand for labor softens. Companies do not need to announce mass layoffs for the shift to matter. They can allow natural attrition to reduce headcount, slow graduate hiring, and rebuild the enterprise around fewer people. The social consequences are slower-moving than those of a recession, but potentially more durable.

Roosevelt would have understood the danger. Capitalism needs workers to believe that effort leads to security, mobility, and participation in national prosperity. If the gains from AI flow overwhelmingly to shareholders while displaced workers are left to retrain for jobs that may not exist, the problem moves beyond economics. It becomes a legitimacy crisis.

The lazy answer is that new jobs always appear. History offers some comfort, but this is far from certain. In prior periods of displacement, new industries had already risen before the displacement began. That is not the case today. New jobs must pay comparable wages, exist where displaced workers live, and be accessible to mid-career professionals whose skills were built around the old model. They must arrive before family savings, mortgages, and retirement plans are damaged beyond repair. Companies cannot be allowed to keep the productivity gains while taxpayers absorb the social cost.

Roosevelt would not have dismissed AI unemployment as a market adjustment. He would have understood it as a threat to public order, household security, and the legitimacy of capitalism.

Federal Oversight

Roosevelt’s approach to corporate power began with an investigation. His presidency created the Bureau of Corporations in 1903 within the new Department of Commerce and Labor, giving Washington a way to examine large interstate companies, understand how they operated, and expose abuses that individual states could not realistically police. This was central to his trust policy. Roosevelt did not want the federal government guessing at the behavior of national corporations. He wanted information, disclosure, inspection, and administrative capacity that matched the scale of the businesses being examined.

This was the meaning of “publicity” in the Progressive Era. Corporate power had to be made visible. Hidden rebates, discriminatory pricing, preferential railroad treatment, opaque ownership structures, and political influence were not just questionable business practices. They were mechanisms through which private scale weakened public authority. The Bureau of Corporations gave Roosevelt a fact base from which to act, while the Hepburn Act of 1906 strengthened oversight of railroad rates by expanding the authority of the Interstate Commerce Commission.

His third-party campaign extended the same philosophy. The New Nationalism speech and the Progressive Party platform called for a stronger national government, fuller disclosure of corporate affairs, labor protections, social insurance, and federal machinery capable of dealing with industrial capitalism on national terms. Roosevelt’s answer to concentrated power was not rhetorical outrage. It was supervision backed by institutional capacity.

Applied to AI, the implication is straightforward. Roosevelt would not have accepted voluntary ethics statements, corporate self-certification, or marketing language about responsible innovation. He would have wanted a federal authority with the power to investigate the firms that build and deploy the technology. Its mandate would include AI-driven layoffs and hiring freezes, labor-market effects, model risk, data use, market concentration, compute ownership, energy demand, government contracts, lobbying, and political influence.

The purpose would be to give the public sector enough visibility to understand where power was accumulating, how labor markets were changing, and whether private incentives were aligned with public costs. AI is moving faster than the institutions built to monitor it. That imbalance is precisely the kind of problem Roosevelt believed the federal government existed to address.

A Square Deal for the Algorithmic Age

A Rooseveltian response to AI would begin with the principle that property rights are not absolute when they collide with human welfare. Roosevelt was no socialist, and he did not treat private ownership as illegitimate. He believed capitalism needed strong companies, investment, and rewards for enterprise. But he also believed property existed within a social framework. When private wealth became powerful enough to damage the public, the state had both the authority and the obligation to intervene.

That is the right foundation for an AI compact. The leading AI platforms should be allowed to grow, innovate, and drive productivity, but their gains cannot be treated as purely private if the consequences fall on workers, families, and taxpayers. If AI raises margins by reducing labor demand, the public has a legitimate claim on how that transition is managed.

Roosevelt did not describe work as a human right in the modern legal sense. His view was more practical and more forceful. Decent work, fair wages, family security, and protection from exploitation were obligations of a democratic society. A capitalist system that generated vast wealth while leaving workers exposed to instability was not merely unequal. It was politically dangerous.

A modern Square Deal would start with mandatory labor-impact disclosure. Large firms deploying AI at scale should report how adoption affects headcount, hiring, graduate recruitment, contractor use, wage bills, and productivity. If investors are going to reward companies for efficiency, the public should understand how that efficiency is being achieved.

A Roosevelt compact should also include wage insurance, portable benefits, and public-service employment in areas where society has urgent needs, including care, education, conservation, infrastructure, and local government capacity. An AI adjustment levy on firms that replace labor at scale while margins expand should also be considered. This is not a punishment for innovation. It is a way to ensure that private automation does not create public costs without contribution.

Competition policy would need to address control of compute, data, cloud infrastructure, distribution, and enterprise access. Public-interest AI infrastructure should be available to schools, hospitals, courts, small businesses, and government services, so the public sector is not permanently dependent on private platforms.

This is not a Federal Reserve problem. If unemployment rises because companies become more efficient, lower rates will not recreate the jobs that productivity gains have eliminated. The response has to be fiscal, institutional, and democratic.

Roosevelt regulated capitalism to preserve its legitimacy. That is the correct frame for AI. The goal is not to stop the algorithmic age, but to ensure it remains compatible with a society in which work still supports dignity, households, and broad participation in national prosperity.

Buy Growth Without You on Amazon

Scott Galloway is a podcaster, entrepreneur, and branding expert. The podcast Pivot, which he co-hosts with technology journalist Kara Swisher, ranks first on Apple Podcasts in the News Commentary category, and while it doesn’t command the audience of Joe Rogan or Smartless, it is highly regarded for its assessments of technology, media, and innovation trends. He is the author of half a dozen books on topics as diverse as the post-COVID economy and the challenges facing young men, for which his writing has received acclaim. Galloway doesn’t hide his political biases and is often critical of the lack of regulation facing the technology sector and its toxic influence on society. Self-deprecating and hysterically funny, he is one of the leading figures in new media and technology.

Like many pundits (including me), Galloway occasionally overreaches by offering opinions on topics where his knowledge base is lacking. Geopolitics is one of those areas. He recently stated that if China wants to hurt the US economy, it could flood the US ecosystem with free, open-source LLMs to completely disrupt the rollout of AI spending, as the debate rages over the long-term viability of hyperscale spending patterns. This does show an ignorance of how Washington has handled Chinese technology threats in the past, but he is hardly alone. We have recently seen a renewed debate about the adoption of China’s large language models and the idea that they are disproving the scaling laws at the heart of the justification for hyperscalers’ estimated $725 billion in spending on data center assets in 2026 alone. The debate is deeply flawed on myriad levels, but the simplest way to discredit the bearish arguments is that, like almost all Chinese technology, Chinese Frontier Models will eventually be banned in the United States.

Many investors believe that the capital being deployed into chips, data centers, power, and cooling has outpaced what near-term use cases can justify. In such an environment, bad news is latched onto with greater vigor than good news, and every data point that supports the excess narrative is treated as more important than the one before it. That is why Z.ai’s GLM-5.2 has reignited the same argument that followed DeepSeek earlier this year. Chinese labs are improving, and any evidence that credible models can be produced or distributed at lower cost will lead investors to question Nvidia’s margins, hyperscaler capex, data-center utilization, and power demand.

Z.ai describes GLM-5.2 as a long-horizon model with a one-million-token context window, while Business Insider reported that it has drawn attention from Silicon Valley developers for coding tasks and agentic workflows. The claims will be debated, as they always are with Chinese technology, but if capability can be delivered with less capital, investors will question the infrastructure bill.

Investors should not dismiss these questions. The correct response to GLM-5.2 and DeepSeek is not to wave away the possibility that model pricing falls faster than expected, but to separate three issues that are being conflated:

1. Do Chinese models undermine the economics of the AI infrastructure buildout?

2. Are Chinese LLMs viable competitors in the United States and Europe?

3. Can U.S. labs replicate the performance and efficiency gains being claimed by DeepSeek, Z.ai, and others?

Subscribe now

Why Chinese LLM Progress is Misleading?

The first question is whether Chinese models undermine the economics of the AI infrastructure buildout. The answer is more nuanced than the market reaction suggests. DeepSeek did not make compute irrelevant, and GLM-5.2 does not prove that U.S. capex is irrational. Both point to the same conclusion: model efficiency is improving quickly. That will pressure pricing, force more discipline around data-center returns, and expose poorly conceived projects. It does not follow that the entire infrastructure cycle has been invalidated.

GLM-5.2 should be treated as a datapoint, not a verdict. Long-context coding performance is useful, but it is not the same as broad enterprise-grade intelligence, and aggressive API pricing tells us very little about the economics behind the release. Pricing does not reveal subsidies, utilization assumptions, engineering expense, failed runs, or whether the economics hold once demand scales. DeepSeek created the same analytical error when investors treated a reported final training run as if it represented the full cost of building a competitive AI lab. Z.ai may be impressive, but it is not a balance sheet for the entire Western AI buildout.

The more useful conclusion is that Chinese releases are forcing investors to separate efficiency from excess. Better models reduce the cost of using AI, which should pressure weak business plans and expose poor capital allocation. That is healthy. It does not eliminate the broad need for chips, power, memory, networking, cooling, and high-quality data. Falling unit costs are part of how adoption broadens, not proof that the physical layer has become unnecessary.

Carlota Perez’s work is useful because it explains why waste and progress often arrive together. In Technological Revolutions and Financial Capital, she argued that major technology waves attract capital before the underlying economics can justify it. Speculative excess is part of the installation phase because it finances infrastructure that later becomes indispensable. Railroads, electrification, telecoms, and the internet all followed this pattern. Many investors lost money, but the installed base survived and became the platform on which the next economy operated. AI will almost certainly rhyme with that history. The mistake is to confuse poor returns for some investors with proof that the infrastructure itself was not needed.

The open-source nature of many Chinese LLMs also limits how durable any efficiency advantage can be. If DeepSeek or Z.ai releases a model that appears to do more with less, U.S. labs can study it, copy the useful parts, and apply those lessons to their own systems. The method does not remain trapped inside China once it is published. It becomes part of the global research base. Chinese labs may force the industry to move faster, but they also provide American companies with a roadmap for lowering their own costs.

The valuation debate ultimately comes down to whether revenue can arrive fast enough to justify the infrastructure bill. Bain estimates that AI will need roughly $2 trillion in annual revenue by 2030 to fund the compute capacity implied by current demand forecasts, yet still projects an $800 billion annual revenue gap even after accounting for AI-related savings. Goldman Sachs estimates roughly $7.6 trillion of AI capex between 2026 and 2031 across compute, data centers, and power. MIT’s GenAI Divide report adds the enterprise reality check: despite $30–40 billion of corporate GenAI investment, 95% of organizations in its study had yet to see measurable P&L returns. These numbers do not prove that AI is a bubble. They show why investors are asking whether the revenue curve can rise quickly enough to keep pace with the spending curve.

My pushback is the calendar. 2030 is a useful checkpoint, but it is too early to be treated as the terminal ROI date for what may be the largest infrastructure buildout in modern economic history. Chips depreciate quickly, but data centers, grid upgrades, power contracts, cooling systems, fiber, and enterprise workflow changes mature on different timetables. By 2030, investors should demand evidence that utilization, inference demand, enterprise adoption, and productivity are moving in the right direction. They should not demand that the entire infrastructure cycle has already paid for itself.

From Glut to Shortage and Back Again

The AI debate lurches from shortage to glut with remarkable speed, but I see very little evidence of the latter today. Google’s decision to cap Meta’s use of Gemini is a useful reminder that frontier compute remains constrained. According to the Financial Times, Google told Meta around March that it could not provide all the Gemini capacity Meta wanted to purchase, disrupting some internal AI projects and forcing Meta to push employees to use tokens more efficiently. This was Meta, one of the largest AI spenders on the planet, being rationed by Google. If excess compute were already available across the frontier model ecosystem, Google would not be limiting a customer of that scale.

The move toward quotas should be viewed constructively. One legitimate criticism of the AI rollout is that usage has often been undisciplined, with companies experimenting before they understand the value of each query or workflow. Quotas impose cost discipline. They force enterprises to distinguish between model usage that creates value and usage that is merely computational noise. That helps LLM providers allocate scarce capacity to higher-value demand and pushes corporates toward measurable use cases rather than indiscriminate consumption.

As we stand today, the world has an intense compute shortage. McKinsey estimates that AI-related data center capacity demand could reach 156 GW by 2030, requiring 125 GW of incremental capacity from 2025 onward and roughly $5.2 trillion of investment. That gap will not be bridged in twelve months because the bottleneck is no longer just GPUs. It is substations, transformers, gas turbines, grid interconnections, permitting, cooling systems, skilled labor, and the ability of U.S. utilities to deliver firm power where the data centers are being built. The IEA expects data centers to account for almost half of U.S. electricity demand growth between now and 2030, which means AI is becoming a power-deployment problem as much as a semiconductor problem.

The realistic window to close the current shortage is probably 2028 to 2031, with the caveat that the United States will increasingly rely on behind-the-meter gas generation and microgrids where grid connections are too slow to come online. The ROI of bridging that gap depends on utilization, token demand, and the speed at which enterprise use moves from experimentation to embedded workflow. If gluts appear, they are unlikely to be broad before 2030. They are more likely to emerge first in the wrong geographies, with weak power arrangements, poor interconnection prospects, or data centers built for demand that never arrives.

It’s a Moot Point: US Companies Don’t Use Chinese Software

The mistake is to assume that cheaper Chinese models will be allowed to derail the U.S. AI sector through broad enterprise adoption. U.S. corporates do not use Chinese-made software outside China unless local operations require it. The corporate stack is Microsoft, Salesforce, ServiceNow, Workday, Oracle, Adobe, AWS, Google, Snowflake, Databricks, and a long list of U.S. and allied vendors. This is procurement discipline, compliance process, intellectual property protection, and board-level risk management. No general counsel wants to explain why confidential contracts, source code, customer records, or internal strategy documents are being processed through software governed by Chinese law.

The application layer reinforces this reality. Western enterprise workflows already sit inside American platforms, and no Chinese AI model can be embedded deeply into Microsoft 365, Salesforce, or Workday in the way that OpenAI, Anthropic, Google, or Meta can be routed through trusted vendors. The line was drawn before the current AI race began. A strong benchmark does not change the procurement, compliance, and national security obstacles that govern corporate software adoption.

The United States will eventually ban all Chinese LLMs.

This is not merely a corporate governance decision. Washington has already decided that Chinese technology products capable of collecting, transmitting, or controlling sensitive data should not be treated as ordinary commercial goods. Congress, regulators, and executive agencies have moved against Chinese telecom equipment, surveillance cameras, drones, social media platforms, and connected vehicles, while export controls have targeted the advanced semiconductors and equipment that would strengthen China’s AI and military capabilities. The legal authorities differ, but the direction is consistent. When a Chinese technology product creates a pathway into American data, the policy response has been restriction, forced separation, or outright prohibition. LLMs sit squarely inside that pattern.

Huawei is the template. ZTE and DJI were already on Washington’s radar, but Huawei became the case study for how the risk posed by Chinese technology would be handled once it entered the national security system. The public never saw the full Huawei file because much of the evidence was classified, which meant the debate outside government was conducted with incomplete information. That ambiguity mattered for public perception, but it did not matter for policy. Once the intelligence community concluded that Huawei could not be treated as separate from the Chinese state, the commercial arguments became secondary.

That conclusion survived Trump 1, Biden, and now Trump again. Chinese technology risk is one of the few bipartisan policy threads in Washington, and the framework has expanded beyond telecom. The FCC’s covered list includes communications equipment and services deemed to pose an unacceptable national security risk, including Huawei and ZTE, with Hytera, Hikvision, and Dahua covered in specified contexts. Commerce has extended similar logic to connected vehicles linked to China and Russia because software control, remote connectivity, and data access have become security concerns.

Washington has already shown a willingness to pressure an American AI company when model access collides with national security demands. Anthropic was reportedly designated a supply-chain risk after a dispute with the Pentagon over military use of Claude, including restrictions related to mass domestic surveillance and fully autonomous weapons. If U.S. officials are prepared to challenge an American model provider operating under American law, Chinese LLMs will face a harsher review when China’s National Intelligence Law requires organizations and citizens to support, assist, and cooperate with state intelligence work.

Apple’s reported interest in Chinese memory chips is another reminder that corporate expediency and national strategy frequently collide. Apple is reportedly seeking U.S. approval to purchase memory from CXMT, a Chinese supplier on a Pentagon blacklist, as AI-driven demand contributes to higher component costs. The procurement logic is obvious, but China hawks will see the episode as confirmation that large companies will use Chinese supply when shortages bite unless Washington prevents them.

The national security concern around DeepSeek is far more severe than the concern around a Huawei smartphone or FEMA using a DJI drone that is not connected to the internet. The Huawei threat was never fully visible to the public because much of the evidence was classified. DeepSeek will be easier to explain because sensitive information can enter the system through the front door. No one needs to break into a network when employees voluntarily place proprietary information into a model governed by Chinese law.

Conclusion: China is an Exaggerated Threat.

The consequence of Chinese LLM progress is, therefore, almost certainly being misread. The lasting impact will not be that DeepSeek, Z.ai, or any other Chinese frontier model becomes the operating layer for corporate America. It will be that U.S. labs study the open-source releases, absorb the useful weights, methods, architecture choices, and efficiency gains, and incorporate them into their own research platforms. That is the real transmission mechanism. If Chinese labs find ways to produce more capability with less compute, the benefit does not remain trapped inside China once the model is released. It becomes part of the global research base, and the companies best positioned to commercialize those lessons in the United States are still OpenAI, Anthropic, Google, Meta, Microsoft, Amazon, SpaceX and the rest of the American AI ecosystem.

Investors can and should be skeptical about the claims attached to Chinese models. Reported training costs may exclude failed runs, subsidies, engineering expenses, infrastructure depreciation, or the true cost of scaling demand. API pricing may be strategic rather than economic. Benchmarks may overstate commercial usefulness. Those are legitimate questions, but they are not the same as saying Chinese LLMs threaten U.S. AI hegemony through direct adoption. The more obvious conclusion is staring us in the face. Nearly every major Chinese technology that creates a pathway into American data, moves that data through Chinese-controlled systems, or stores it on servers outside U.S. jurisdiction has eventually been treated as a national security risk. Telecom equipment, surveillance cameras, drones, social platforms, connected vehicles, and now frontier AI all fit into the same framework. There are exceptions at the margin, but the direction of policy has been unmistakable.

That is why Chinese LLMs should not be used as a discounting mechanism for the business models of Gemini, Anthropic, OpenAI, or the broader U.S. AI stack. They may pressure pricing. They may force faster efficiency gains. They may expose weak infrastructure projects and unrealistic return assumptions. But the competitive threat is indirect. Chinese labs can accelerate the research frontier; they are unlikely to own the enterprise layer in the United States. The real lesson from DeepSeek and GLM-5.2 is that American labs may be able to do more with less as they absorb insights from Chinese models. That strengthens the U.S. AI ecosystem more than it weakens it. To fear Chinese LLMs as direct competitors inside corporate America is to lose sight of the obvious: the models may travel through open weights and research papers, but the products will not be allowed to travel through American enterprise infrastructure.

AI did not start this process. It is completing it.

Buy Growth Without You on Amazon

The Rhythm That Faded

Growth Without You is a book about a profound economic transition hiding in plain sight. For most of the twentieth century, investors, policymakers, businesses, and workers organized their expectations around the business cycle. Economies expanded, overheated, contracted, shed labor, reset excesses, and began again. Recessions were not accidents. They were the organizing rhythm of post-World War II capitalism. Portfolio construction, central banking, fiscal policy, career planning, and retirement assumptions were all built on the belief that this rhythm was permanent.

This book argues that the rhythm has faded. Not because governments became brilliant, or because central banks learned how to abolish recessions, but because the economy itself changed. The old cycle was a function of friction: slow information, rigid supply chains, poor inventory visibility, expensive coordination, labor immobility, and capital that could not adjust quickly enough to prevent small imbalances from becoming large ones. Over the past fifty years, containerization, enterprise software, globalization, the internet, mobile computing, and real-time data dismantled those frictions layer by layer. Companies became more efficient, more responsive, and less prone to the errors that once made downturns self-reinforcing. By the time artificial intelligence arrived, the traditional business cycle had already been weakened. AI did not start this process. It is completing it.

The Final Compression

The core argument of Growth Without You is that artificial intelligence represents the final compression of economic friction: the automation of cognition itself. Prior technologies improved how information moved. AI changes who interprets it, who acts on it, and how much human labor is required to turn knowledge into output. It is not merely a technology-sector story. AI embeds itself in workflows across every sector of the global economy. Its economic significance lies in its ability to reduce the cost of judgment, coordination, compliance, analysis, diagnosis, scheduling, and decision-making across the entire enterprise.

The Margin Dividend

The result is a new phase of corporate margin expansion. The first wave of the AI trade rewarded infrastructure: semiconductors, data centers, cloud providers, power generation, and the model builders. The next wave is broader and more important. It belongs to ordinary companies that use AI to remove administrative cost, improve asset utilization, automate customer service, compress compliance burdens, reduce error rates, and scale output without rebuilding headcount. A business already earning 30% margins does not transform itself by saving a few percentage points of cost. A business earning 5% to 8% margins can change its entire equity story by adding three durable points of profitability. That is the margin convergence at the heart of the book. The AI dividend will not remain trapped in Silicon Valley. It will spread across sectors and geographies as low-margin businesses gain access to tools once available only to the most advanced technology platforms.

Growth Without the Worker

But the same force that lifts margins weakens labor, which is the uncomfortable tension in the title. Growth Without You is not an anti-AI book. It is an assessment of who captures the gains. For most of modern economic history, growth required more workers. That relationship is breaking. AI allows firms to expand output without proportional employment growth, raising productivity, revenue per employee, margins, and free cash flow while reducing the need for broad-based hiring. The adjustment is unlikely to arrive as one dramatic unemployment shock. It is more likely to emerge through narrower hiring funnels, fewer entry-level roles, smaller graduate intakes, thinner middle-management layers, automated back-office functions, slower wage growth, and companies learning that they can grow without adding people at the historical rate.

That has enormous implications for economic activity. Productivity statistics may look better while the lived experience of many workers deteriorates. The path to 10% unemployment by 2035 does not require a traditional recession. It can come from a decade-long erosion in labor demand as firms replace hiring with software, workflows, and models that do not require salaries, benefits, training, or promotion. Corporate profits can surge, asset prices can compound, and the economy can continue expanding, even as the household sector begins to fray because the wage income supporting consumption weakens. That is the central paradox of the AI economy: growth continues, but it becomes less labor-intensive; profits rise, but the consumption base weakens; capital wins, while the employment model that supports demand comes under pressure.

Capital wins, while the employment model that supports demand comes under pressure.

When the Old Tools Fail

The consequences for monetary and fiscal policy are equally profound. Central banks were built to manage an economy where labor markets, wages, credit creation, and inflation moved in recognizable cycles. That economy is disappearing. If AI allows companies to protect margins while slowing hiring, the old relationship between unemployment, wages, and inflation becomes less reliable. If employment lags output, central banks may find themselves using twentieth-century tools on a twenty-first-century production function. The book argues that the Federal Reserve’s experience from 2022 to 2024, when the sharpest tightening cycle in forty years failed to produce the recession most models expected, was not an anomaly. It was evidence that the economy those models describe no longer exists.

Fiscal policy faces an even harder problem. Governments tax labor income because historically, labor has generated the bulk of the economic gains. But if AI shifts more income toward profits, dividends, capital gains, intellectual property, and platform ownership, then the tax base becomes much less stable. Social safety nets built around employment income become structurally strained. The political compact between growth and shared prosperity weakens. Governments will be forced to choose between taxing capital more aggressively, redesigning social insurance, accepting wider inequality, or pretending the old framework still works long after the evidence says otherwise.

The Mismatches

Growth Without You is ultimately a book about a series of mismatches: between the economy we still analyze and the economy now emerging; between the tools policymakers use and the mechanisms that drive outcomes; between how investors allocate capital and how wealth is created; and between how workers plan their lives and how the labor market will reward them. It does not argue that the future is dystopian; the future is being misread. AI will make companies more efficient, margins higher, capital more powerful, and traditional cycles less relevant. The question is whether individuals, investors, and governments adapt before the mismatch becomes impossible to ignore.

About the Author

Paul Krake is an investor, strategist, and independent researcher with three decades at the intersection of macroeconomics, policy, geopolitics, technology, and capital markets. In 2011, he founded View from the Peak, a multi-asset research platform that advises endowments, family offices, asset managers, hedge funds, and pensions across twelve countries.

Before that, he spent 18 years in investment banking and asset management, including as Co-Managing Partner of the New York hedge fund Corus Capital and in Asia-focused roles at Caxton Associates, Moore Capital, Goldman Sachs, and Macquarie Bank. He holds a degree in economics and politics from Monash University in Melbourne.

Subscribe now

Subscribe now

The headline from SpaceX’s listing was the $75 billion it raised, the largest IPO ever recorded. The proceeds are the least interesting part of the story.

What Musk acquired when SpaceX hit the Nasdaq on June 12 was not cash. It was currency. The stock priced at $135, closed its first day near $161, and ran higher again the following session; the company crossed a $2 trillion market value almost at once. Four days later, SpaceX signed a definitive agreement to buy Anysphere, the maker of the coding tool Cursor, in an all-stock merger valuing the startup at $60 billion. No money changes hands. Cursor’s investors receive SpaceX shares, converted at a seven-day average price just before the deal closes in the third quarter.

Then the market did something worth sitting with. SpaceX stock rose about 11% when the agreement was confirmed. A Cursor purchase dilutes existing holders by roughly 3.5%, yet at a $2 trillion base, a single session added more than $200 billion of market value, over three times the $60 billion price. The purchase did not cost SpaceX shareholders. It paid them.

That is the part to understand. The binding constraint on any all-stock acquisition is the quality of the paper. Issue cheap stock to buy a company and you punish your owners through dilution. Issue pricey stock and the math inverts: the more the market prizes your shares, the less of the company you surrender to acquire someone else’s. Musk now holds the most expensive script on the market, and he can overpay for almost anything and let the share price carry the bill.

This is the next phase of AI funding maturation. Initially, it was about raising money to rent compute from VCs and hyperscaler partners. This is evolving into the public market domain and is about minting a currency expensive enough to buy capability, distribution, and finished product without spending a dollar of it. Cursor is the proof of concept, and a telling choice: SpaceX did not buy a frontier model it could have built inside xAI. It bought the surface where millions of engineers already work. The same logic reaches any application layer, any data business, any distribution channel that would widen the moat.

Set that against the position of the pure model companies. Anthropic and OpenAI are each laying groundwork for their own listings, and that will help them with their own form of currency. However, they will still sit a tier below the hyperscalers, and now below SpaceX. The disadvantage is one of position, not effort. Amazon funds Anthropic and sells it chips; Microsoft funds OpenAI and resells its models. The hand that pays for the compute also competes for the customer. A model company rents its scarcest input from the firms it is trying to beat. The hyperscalers are both benefactors and overwhelming competitors to the model companies that are increasingly looking like one trick ponies.

I remain skeptical that the standalone model businesses ever reach durable profitability. The product is commoditizing. Frontier capability is converging and switching costs stay low. Price becomes the field of battle. When the output commoditizes, margin migrates to whatever stays scarce, and what stays scarce is the compute. The infrastructure layer absorbs the economics; the model layer passes them through. A lab can win every benchmark and still hand its margin upstream to the data center that trained it.

The winners are the houses that own more than a model. The hyperscalers profit at three points in the same transaction: they sell the compute, they take their cut of the global rollout, and they fold the best tools into products that already reach billions of people. SpaceX has just bought a seat at that table. Starlink revenue and the xAI compute base sit under one balance sheet, alongside the Grok models and, now, Cursor’s distribution to working engineers. The whole apparatus is financed by stock the public market has chosen to value at more than 100 times sales.

None of this asks the rocket business to justify the price. None of it demands answers about how a data center in orbit actually works. The market has decided to treat SpaceX as a hyperscaler even if It has revenue 6% of that of Microsoft. Musk has told shareholders he believes the firm can approach $1 trillion in revenue by 2030, against roughly $19 billion last year. Whether that number arrives is almost beside the point today. The valuation is already doing the work the cash flow has not yet earned.

A year ago, the contest looked like two well-funded labs against three hyperscalers, with SpaceX nowhere on the board. It is now a tier-one competitor backed by investors prepared to underwrite one of the most expensive companies in living memory.

Subscribe now

My twelve years in New York were the best of my adult life.

For someone who loves sport, New York is an extraordinary place to live. That said, I didn’t have much luck with the teams I chose to support. My New York Jets are, arguably, the worst franchise in the NFL over the past 30 years and, frankly, the laughingstock of the league. Yes, I am a New York Yankees fan and have seen them win multiple World Series, but I have also been a diehard New York Knicks supporter throughout that entire period. I saw them lose the Finals in 1999. Being a Knicks fan over the past quarter of a century has involved a lot more emotional scar tissue than joy.

Which is why my jealousy today is palpable.

New York is joyous in a way that only New York can be. I have dear friends who are now watching these games with their adult children, people I grew up with in the city, and the happiness is endless. There is something uniquely powerful about a city rediscovering a team that has always belonged to it emotionally, even when the results have provided little justification for the loyalty. The Knicks are not just a basketball team. They are a civic institution, a shared language, a memory bank, and, for many New Yorkers, an inherited form of suffering.

Where else could you watch Larry David and John McEnroe, arguably two of the grumpiest men of all time, hugging at the end of the game. New York sport can do that. It can temporarily disarm even the most committed curmudgeons.

That said, if I were in New York this week, it would be a little disappointing knowing that I probably could not afford to go and watch live.

The ticket prices at Madison Square Garden have become as much a story as the basketball itself. This isn’t a standard premium for a major sporting event. We are talking about prices that are multiples of what people pay to attend the Super Bowl or the World Cup final. There are seats changing hands at levels that sit somewhere between absurd and revealing. When people are paying $50,000 to $75,000 to watch a 48-minute basketball game, the price is no longer just a function of sporting demand. It is an economic signal.

New York City’s per-capita income is a little over $50,000 a year. For many seats in Madison Square Garden this week, that is the price of admission. That comparison is stark. One year of income for the average New Yorker converted into one night of access for someone else.

The New York Knicks are the epitome of the K-shaped economy. On one branch are the people who own appreciating assets: businesses, equities, carried interest, private companies, real estate, intellectual property, and the platforms that dominate the modern economy. On the other branch are the people who rely on wages, rent-burdened incomes, fragile employment, and the promise that hard work will eventually translate into security. Both groups may love the Knicks. Both groups may feel the same emotional release when the final buzzer sounds. But they are not participating in the same economy.

There have been endless, and endlessly wrong, proclamations of New York’s demise. People have been predicting the death of New York City for as long as I can remember. Crime, taxes, remote work, commercial real estate, progressive politics, the pandemic, the cost of living, and now the mayorship of Zohran Mamdani have all been presented as reasons why capital will flee and New York will finally lose its status as the world’s most important city. Perhaps New York will always give its critics enough ammunition to keep the argument alive. It is expensive, badly governed, unequal, chaotic, and often maddening.

But look at Madison Square Garden.

If last night’s snapshot into the affluence of New York City is any guide, the proclamations of death are grossly exaggerated. Panning through the crowd on TV at three o’clock in the morning, London time, and seeing hedge fund luminaries, venture capital executives, Hollywood celebrities, private equity billionaires, and generational family offices littering the arena’s front rows gives you pause. This is a city of extraordinary, entrenched wealth, and there is true value in living at the epi-center of global wealth creation.

But it also reveals something uncomfortable.

The joy belongs to the city. The seats belong to the capital.

Artificial intelligence is the K-Shaped Economy on steroids.

Another way to think about the 20,000-odd people who filled Madison Square Garden to the brim last night is not simply through their common love of the New York Knicks, but through their collective ownership of the companies and assets that will drive artificial intelligence. When thinking about the ownership of capital, there was no better example than that audience. This crowd has been, and will continue to be, the beneficiary of capital’s rise and labor’s decline over the past forty years. Whether through company ownership, entrepreneurship, public equity portfolios, private-market access, and, increasingly over time, inheritance, those who own and can access capital will be able to invest in AI and the companies that benefit from it. Everyone else will be left watching the party from afar.

This is not a critique of successful people. It is an observation of the inevitable consequences of wealth concentration, driven by capital’s ability to generate superior long-term wealth relative to labor.

If I am right and the United States reaches 10% unemployment by 2035, I have very little doubt that the composition of Madison Square Garden won’t change much at all. I do not see a dramatic fall in ticket prices if the US economy grows structurally more slowly due to slower consumption and a greater fiscal burden. The more important point is that companies will be infinitely more profitable. Profit margins will be significantly higher than they are today. The owners of capital will have benefited from the greatest friction-removal device global companies have ever seen.

So, if the Knicks ever make another run like this again, the Garden may look remarkably familiar. The same class of people will be there: the owners of capital, the deployers of AI tools, the beneficiaries of automation, and the people whose wealth compounds as labor’s bargaining power declines.

The economy outside the Garden, however, will be vastly different.

To understand what is happening to hyperscaler valuations, and why it matters so much for how investors should position themselves, it helps to start with a concept from academic finance that rarely makes it into investment conversations:

The Lucas Tree

Robert Lucas introduced it in 1978 while an economist at the University of Chicago, as a framework for thinking about asset prices in an economy where output falls from a tree, with no investment required to produce it. The owner of the tree collects the fruit year after year. There is no new net investment. The asset generates cash flows in perpetuity with no need to reinvest in the productive base. In this world, the earnings yield and the free cash flow yield are identical, because there is nothing between them. Every dollar of earnings is available to be paid out to owners.

The Federal Reserve Bank of Minneapolis, in a January 2026 paper titled "A Macroeconomic Perspective on Stock Market Valuation Ratios," demonstrated with unusual precision that this relationship has been the dominant driver of corporate valuations for the past forty years. It uses this model as one end of a theoretical spectrum. At the other end sits the standard growth model, in which firms must invest in physical capital to produce any output. In that world, earnings and free cash flow diverge by exactly the amount of net investment, and the earnings yield exceeds the free cash flow yield because a significant portion of what a firm earns must be reinvested to maintain a competitive position. The paper’s central empirical finding is that the U.S. corporate sector, over the past forty years, has been moving steadily from the growth-model end of that spectrum toward the Lucas-tree end. Net investment as a share of enterprise value has fallen dramatically since 1980. The gap between the earnings yield and the free cash flow yield has narrowed from nearly 6% in 1980 to roughly 1% by 2022. Companies have been generating more and more of their earnings as distributable cash, with less and less of it consumed by reinvestment.

For the hyperscalers, this transition was not gradual. It was the entire basis of their valuation story. Google generates revenue from search advertising at essentially zero marginal cost per additional query. Meta’s social platforms scale to billions of users without building a factory for each new market. Microsoft’s software licenses replicate at zero cost. These businesses were, for the better part of two decades, almost perfect Lucas tree economies. They produced free cash flow at extraordinary rates relative to their asset bases, which is precisely why their earnings yields looked low relative to historical norms while their free cash flow yields remained close to the market average. Investors who understood this, and the Minneapolis Fed paper provides the clearest academic articulation of why they were right to understand it, knew that the earnings yield comparison to a manufacturing company was meaningless. The right metric was free cash flow yield, and on that basis, the hyperscalers were not expensive. They were priced at roughly the same cash yield as the broader market, for businesses with structurally higher growth rates and lower reinvestment requirements.

That argument is now under significant pressure, and the pressure comes from the hyperscalers themselves. In 2024, Alphabet, Amazon, Meta, and Microsoft collectively spent approximately $229 billion on capital expenditures, which accounted for roughly 51% of their combined operating cash flow. In 2025, those numbers moved dramatically further.

• Alphabet spent $91.4 billion on capex against an operating cash flow of $164.7 billion, a 55% capex-to-operating cash flow ratio.

• Amazon spent $131.8 billion against an operating cash flow of $139.5 billion, a ratio of 94%, leaving free cash flow of just $7.7 billion for the year, down from $38.2 billion the prior year.

• Meta spent approximately $72 billion against an operating cash flow of $115.8 billion.

• Microsoft spent $64.6 billion against an operating cash flow of $136.2 billion.

In aggregate across the four companies, capital expenditure in 2025 reached approximately $360 billion, against a combined operating cash flow of $556 billion, a ratio of 65 %. Combined free cash flow for the four was $196 billion, meaning capex consumed approximately 183% of their aggregate free cash flow. For Amazon alone, the free cash flow yield on its enterprise value has essentially collapsed to zero.

The valuation implications of this transition are straightforward and follow directly from the Minneapolis Fed framework. As the hyperscalers move from Lucas tree businesses toward growth model businesses, the gap between their earnings yield and their free cash flow yield will widen again, in the opposite direction from the past two decades. Investors who valued these companies primarily on earnings multiples could continue to do so and reach a superficially reasonable conclusion. Investors who valued them on free cash flow yield, which the Minneapolis Fed demonstrates is the theoretically correct valuation metric, now face a substantially different picture. The free cash flow available to equity holders is being consumed by data center construction, server procurement, energy infrastructure, and custom silicon development at a rate that, if sustained, redefines what these businesses are.

They are no longer fruit trees.

The investment conclusion that follows from this analysis is not that the hyperscalers are bad businesses. They are extraordinary. The question is whether their valuations still reflect the Lucas tree characteristics that justified premium multiples when they were capex-light, or whether those multiples need reassessment given the current state of their capital accounts. A sector spending 65 cents of every dollar of operating cash on fixed assets, and guiding toward spending of close to $700 billion in aggregate in 2026, is not the same sector that was priced at 25 to 30 times earnings a decade ago on the assumption that capital investment would remain minimal. The multiple should reflect the reinvestment burden. To the extent that it does not yet fully do so, the valuation gap between the hyperscalers and the rest of the market is overstated, and the relative opportunity in businesses poised to experience capex-light margin expansion for the first time looks even more compelling.

The investment implication follows directly. The companies bearing the capital cost of the AI infrastructure buildout and the companies reaping the efficiency dividend from its consumption are not the same. They are different sectors of the economy and different stages of the margin expansion cycle. Technology had its moment. The capex-light free cash flow machine that drove the S&P 500 for two decades is now deploying capital at rates that would have been unrecognizable five years ago. The beneficiaries of that capital deployment, the businesses across healthcare, industrials, financial services, logistics, and consumer sectors that will access AI-driven efficiency without bearing the investment cost, are beginning a margin expansion journey that technology completed long ago.

This is why the margin gap between technology firms and the rest of the global business universe is not a permanent feature of the economic landscape. It will narrow, though not because technology margins will collapse or because other sectors will achieve anything approaching the returns that software businesses generated at their peak. The compression works from both directions simultaneously. Technology operating margins, which ran at 27% to 28 % through most of the past decade on the back of capex-light business models, are already coming under pressure as the hyperscalers commit hundreds of billions of dollars annually to physical infrastructure. Higher depreciation charges, rising energy costs, and the operational expense of running data centers at scale are beginning to work their way through the income statements of the companies that built the margin premium in the first place.

Meanwhile, the rest of the S&P 500 has barely moved over the past 15 years, stuck in the 11%-13% operating margin range, churning through economic cycles without accessing the operating leverage that technology monopolies have achieved over the same period. Healthcare systems operate at 6% margins. Industrial companies sit at 10 %. Small caps across all sectors are at 4 %. These businesses will not reach 25 % margins. They do not need to. A sustained improvement of 3% to 5 % points over a decade, driven by gradual AI-enabled cost reductions, would represent a transformation in their profitability and a re-rating of their enterprise value.

The gap narrows as technology is pulled back toward capital intensity, while everyone else is slowly lifted by access to tools that reduce the cost of doing business. That is what margin convergence looks like in practice, and it is a very different story from the one the last twenty years have told.

Subscribe now

There is a test every economics book eventually faces: whether anyone still thinks about it ten years after publication. Thinking, Fast and Slow by Daniel Kahneman instantly comes to mind. The Rise and Fall of American Growth, by Northwestern University Economist Robert Gordon, came out in 2016, and I have been thinking about it, arguing with it, and borrowing from it almost continuously since. My copy is covered in highlights, post-it notes, and handwritten bullet points in the margins. It has sat within reach of my desk for the past nine years, and it is the book I return to most often when I am trying to think clearly about where growth comes from and where it is going.

Gordon’s thesis is deceptively simple. The hundred years between 1870 and 1970 were a “Special Century”. Electricity, the internal combustion engine, indoor plumbing, antibiotics, aircraft, and mass communication did not arrive as a procession of incremental improvements. They arrived together, they rewired every sector of economic life, and they could only happen once. You can electrify a country once. You can move households from outhouses to indoor plumbing once, thereby dramatically reducing infant mortality through improved sanitation. Women can enter the workforce en masse, or high school graduation rates can encompass the majority of teenagers, only once. Gordon argues that productivity growth slowed after 1970, not because human ingenuity flagged, but because the scope for that particular kind of transformation had closed.

What makes the book so effective is the detail. Gordon has an almost obsessive commitment to the texture of ordinary life. He tells you what a pound of nails costs, how long laundry took, or the lack of innovation in air travel since the 1950’s. This is not the prose of Robert Caro. The argument does not need it. You are reading something closer to an excavation, and the weight of the evidence is what eventually fascinates you.

The argument then moves to the post-1970 period, and here Gordon is careful rather than dismissive. In his reading, digital technology transformed one of his six dimensions of economic life (information and entertainment) while leaving the other five more or less untouched. The smartphone is extraordinary, but it did not change how buildings are built, how food reaches the table, or how likely a cancer diagnosis is to prove fatal. His test is in scope. By that standard, the internet revolution did not quite clear the bar, and the 1995 to 2004 productivity revival was, on his reading, a genuine ten-year episode driven by a one-time collapse in computing costs rather than the start of a new trajectory.

Gordon’s framework rests on a specific lens: he measures progress through its observable impact on output and productivity growth, with the important qualification that he is acutely aware of how much the official statistics understated the gains of the Special Century. He is eloquent about everything GDP failed to capture in that era, the time saved, the deaths prevented, the drudgery eliminated, the sanitation that showed up as a pipe installation rather than as the value of not dying from cholera. His argument that the statistics understated the Second Industrial Revolution is one of the most compelling sections in the book.

Where my own framework diverges is not on the question of measurement, but on the nature of what was being delivered. In my upcoming book, Growth Without You, I argue that the gains from globalization, software, and digitization after 1980 operated through a different mechanism than the great physical inventions Gordon chronicles. They were not primarily about raising output the way electrification or the internal combustion engine did. They were about cost, efficiency, and the reduction of friction: the collapse in the cost of coordination, the elimination of entire categories of administrative overhead, the compression of supply chains, the democratization of access to markets and information that had previously required expensive physical intermediaries. Gordon measures the post-1970 era and finds the productivity gains modest. I would suggest that this is partly because the gains were accruing in ways that GDP was never designed to capture, and that the distinction between output-raising and friction-reducing innovation is the vital conceptual fault line separating the two frameworks.

That distinction matters enormously for how we assess artificial intelligence. If AI is primarily an output-raising technology, Gordon’s skepticism about whether it can match the Special Century is a serious challenge. If it is primarily a friction-reducing and cost-compressing technology, operating through the same mechanism as the software and digitisation wave that preceded it but at far greater scale and depth, then the relevant question is not whether it shows up cleanly in measured productivity, but whether it is silently reshaping the cost base of the global economy in ways that will compound for decades before the statistics catch up.

Gordon’s colleague at Northwestern, Joel Mokyr, has been pushing back on the broader pessimism for years. The two are publicly known as an economic odd couple, Mokyr calling Gordon, with some affection, his esteemed and much-misguided colleague. Their disagreement is mostly about whether the pipeline of future invention will look anything like the past, but it sits on the same general terrain.

The deeper question, which is the one I spend most of Growth Without You trying to answer, is whether AI changes the category rather than the degree. Gordon’s test is whether a technology spreads across multiple domains of economic life simultaneously. Every digital wave before this one improved how humans processed information, but left the human at the center of every judgment call. AI is beginning to make judgments on its own. Once that line is crossed, the discussion shifts. It is no longer about whether workers become more productive. It is about how much of the work still requires them. That is, on any honest reading, the kind of pervasive structural change Gordon’s own framework was built to identify.

None of which should be mistaken for dismissal. Gordon is right that impressive technology and transformative growth are not the same thing. He is right that the post-1970 pattern has been one of margin expansion rather than output expansion, and that the headwinds of demographics, inequality, and education compound the problem. Where I part company is on whether AI ultimately extends the third industrial revolution or begins a fourth. I think it begins a fourth. I could be wrong about that. Professor Gordon could well be proven correct, again. The next twenty years will settle it.

I have never met Professor Gordon. I have so many things I want to ask him. If I get to say thank you, that would be enough. His work was the direct inspiration for writing Growth Without You. I wanted to engage seriously with the best version of the argument I was pushing against, and there is no better version than his. If I can produce something that causes a single reader to reach for a highlighter, that will be genuinely humbling.

Subscribe now

On February 1, 1979, the Ayatollah Khomeini stepped off a plane in Tehran after fourteen years in exile and went straight to a cemetery. Behesht-e Zahra, the martyrs’ cemetery, was his first stop on Iranian soil, and the speech he gave there was not a call for reconciliation. He announced, plainly, that the Bakhtiar government was illegitimate and that he intended to appoint his own. The United States embassy and the CIA sent observers. They did not bring a Farsi speaker. Washington would not know what Khomeini had said for several days. By the time they found out, American officials were still debating whether this was the Iranian Gandhi arriving to bless a democratic transition. Arguably, the most consequential speech on Iranian soil in 40-plus years had no translation.

Scott Anderson’s King of Kings is a natural companion to Edward Luce’s Zbig, which I recently reviewed here. But where Luce’s portrait of Brzezinski is largely admiring, a strategist of genuine brilliance operating at the edge of American power, Anderson’s book offers no such comfort. Nobody comes out of this account looking good. Not Brzezinski, as the head of the NSC, not Secretary of State Cyrus Vance, not Ambassador William Sullivan, and emphatically not Jimmy Carter. Anderson’s deeper argument is not simply that the United States mishandled the Iranian revolution. It is that the mishandling was not an aberration. It was the pattern, running from the 1953 CIA coup that restored the Shah through the hostage crisis and on into every subsequent decade. American policy toward Iran has been less a strategy than a recurring failure, dressed in different clothes each time. That this highly readable, eerily relevant, and thoroughly enjoyable book lands in 2025 is timing Anderson could not have planned but could not have improved on.

The Carter administration’s particular incompetence was to compound the usual American vices, overconfidence, misreading of local forces, projection of Cold War anxieties onto situations they did not fit, with an additional dysfunction of its own. Washington did not suffer from a lack of plans in 1978 and 1979. It suffered from an excess of incompatible ones. Brzezinski’s NSC was cabling Sullivan in early November 1978 to tell the Shah that the United States supported him without reservation, while Sullivan himself had already sent his “Thinking the Unthinkable” cable to the State Department arguing that the Shah’s social base had eroded past recovery and that the administration needed to plan for life after him. Two separate scripts, circulating simultaneously, from two arms of the same government. Nobody talking.

Brzezinski read Iran as a geopolitical test. The Shah was a pillar of regional order and Soviet containment, and abandoning him would signal weakness and open space for Moscow. That instinct was wrong in ways that were both clear and consequential. Khomeini’s movement was Islamist, fiercely anti-communist, and determined above all else to be independent. The Soviet specter was mostly a Washington projection, a Cold War reflex applied to a revolution that had its own logic and owed nothing to Moscow. Vance’s State Department was closer to reality in reading the monarchy as finished, but its alternative, managed transition and political accommodation, depended on a window that had already closed, and on a Khomeini that did not exist. Henry Precht, the State Department’s senior Iran hand, was operating on a version of Khomeini as a detached moral nationalist rather than a revolutionary claimant. It was Precht who coined the comparison, calling Khomeini the Gandhi of Iran. The cemetery speech, had anyone understood it in real time, would have corrected that immediately.

The comedy in King of Kings, and Anderson is very good at bureaucratic comedy, is the comedy of paper generating faster than comprehension. Cables crossing cables. Envoys shadowing envoys. General Huyser’s mission to Tehran was so ambiguous in its mandate that some read it as preventing a military coup, while others saw it as preparing one. Officials who were supposed to be coordinating were using different maps because the NSC and the State Department had effectively stopped sharing a single map. This was not a communication failure in the ordinary sense. It was two institutional cultures, with different theories of the world, running parallel policies through the same crisis.

As for the Shah himself. Mohammad Reza Pahlavi emerges from Anderson’s account as a man whose ego, vanity, and breathtaking hubris made Carter look decisive by comparison. Carter at least had the intellectual capacity to grasp what was happening around him, even if he lacked the will to act on it. The Shah had neither. The obvious question Anderson keeps returning to is why the Shah waited so long, dithering through months of escalating crisis when it was apparent to almost every outside observer that his regime was finished. The answer, in Anderson’s telling, is a classic pathology of absolute power. Nobody around the Shah was prepared to tell the King of Kings what was happening in his own country. Sycophants schmoozed, generals dissembled, and ministers managed upward. The Shah existed inside a closed loop of his own myth, and by the time reality forced its way in, the window for any meaningful response had long since closed. It is a dynamic familiar to students of autocracy. It is also, Anderson implies, a dynamic the Americans enabled and then failed to see past themselves.

Carter sits at the center of this, and Anderson’s portrait is not flattering. The president’s failure was not choosing the wrong plan. It was declining to choose at all. He let Brzezinski and Vance run their competing lines, distracted himself with other foreign policy initiatives, and repeatedly arrived late to the scale of what was happening in Iran. By the time decisions were forced, they had become acknowledgments of fait accompli rather than acts of leadership. The sharpest illustration is the decision to admit the Shah to the United States in October 1979, against the explicit anxiety of almost everyone who understood the embassy’s exposure. Carter asked his advisers what they would say when the embassy was overrun and Americans taken hostage, received silence in reply, and went ahead anyway. The embassy was seized on November 4, 1979. Fifty-two Americans were held for 444 days.

That scene is the book’s moral center. Carter saw the risk, articulated it out loud, and still drifted into the choice most likely to provoke it. Anderson’s point is not that one better decision in 1979 would have fixed the relationship. It is that the United States has been making the same category of errors, misreading ideology, exporting its own anxieties, and substituting bureaucratic activity for strategic clarity, since Eisenhower signed off on the coup in 1953. The names change. Today, the dysfunction is particularly consistent.

Subscribe now

Subscribe now

In 1997, Zbigniew Brzezinski published The Grand Chessboard: American Primacy and Its Geostrategic Imperatives. In it, he wrote: “Potentially the most dangerous scenario would be a grand coalition of China, Russia, and perhaps Iran, an anti-hegemonic coalition united not by ideology but by complementary grievances. It would be reminiscent in scale and scope of the challenge once posed by the Sino-Soviet bloc, though this time China would likely be the leader and Russia the follower.” Written without the knowledge of Putin, Xi Jinping, or Donald Trump, it is one of the more useful lenses for understanding what is happening in the world today.

His books rarely exceeded 250 pages. He published 22 of them. Henry Kissinger’s books were the size of house bricks, generally to fit in all the commentary about his own importance to world affairs. This contrast tells you a lot about the two most consequential American foreign policy minds of the postwar era. Both European-born émigrés had thick accents, which created distrust among Washington’s WASPs in the 1960s and 1970s. Both academics who craved power. Both were convinced that ideas about world order mattered more than almost anything else. Kissinger was German-Jewish, Nixon’s man, the architect of détente. Brzezinski was a Polish Catholic, Carter’s man, a Democrat who believed the Soviet Union was not a permanent feature of the international system to be managed but a fragile empire to be broken. One of them was right.

Edward Luce’s Zbig is the biography this comparison deserved. What emerges from Luce, the Financial Times‘s chief US commentator, is a portrait of someone whose views on Russia were never abstract. His father was posted to Canada in 1938 as the Polish consul general in Montreal, and while he never saw firsthand his country being consumed first by the Nazis and then by Stalin’s Red Army, it defined him. His McGill master’s thesis predicted the eventual disintegration of the Soviet Union along nationalist lines. You don’t need a psychological theory to explain Brzezinski’s worldview. You need to know how he grew up.

He took his PhD at Harvard in 1953 and joined the faculty there. After being denied tenure in 1960, he moved to Columbia, where he directed the Research Institute on Communist Affairs and spent the next two decades building the career that would take him to the Carter White House as National Security Adviser in 1977. Kissinger had obtained tenure at Harvard before him, and the national security adviser role before him, too. The day Brzezinski learned of Kissinger’s 1969 appointment, he bought a notebook to record the names of people he would hire when his own turn came.

The Kissinger comparison runs through Luce’s book because it ran through Brzezinski’s life. They insulted each other publicly and regularly. Brzezinski attacked Kissinger’s détente policy as “amoral” and dismissed his approach as “stuntman acrobatics.” Kissinger, in a 1976 Ford administration meeting later declassified, said: “Brzezinski is a total whore. He’s been on every side of every argument.” When Brzezinski was shown the quote in 2012, he replied: “Henry is a friend of mine. He must have meant ‘bore.’” That said, they had lunch together regularly. When Brzezinski died, Kissinger wrote: “The world is an emptier place without Zbig pushing the limits of his insights.” These two men made each other better and more annoying in roughly equal measure. Kissinger told Luce that Brzezinski ranked among the top two national security advisers for strategic thinking. The other one he presumably had in mind was himself.

The policy differences were real. Kissinger looked at the Soviet Union and saw something to be managed. Brzezinski looked at it and saw something to be broken, specifically by targeting the nationalism of the captive states, Ukraine, Poland, the Baltic republics, the peoples the Soviet system had never actually absorbed.

Who was the hawk?

His deep friendship with Karol Józef Wojtyła, who became Pope John Paul II in October 1978, was part of that strategy. The Pope’s moral authority in Poland and his connection to the Trade Union, Solidarity, and the Polish democracy movement were central to the 1989 collapse of the Soviet Union. The accusation that Brzezinski had rigged the papal conclave was one of many conspiracy theories thrown his way across a career that generated enemies with remarkable consistency. He was accused of antisemitism for criticizing Israel’s Washington lobby and Israel’s role in undermining the Camp David Accords. A national hero in Poland, he was accused of betrayal by the Polish right when he refused to endorse the theory that the 2010 Smolensk air crash that killed President Lech Kaczyński and 95 others was Russian sabotage. He accumulated accusations the way people with profound conviction and moral strength do.

The Carter years were his moment of greatest power and greatest frustration. Carter was a decent man whose instinct for consensus made Brzezinski’s job enormously difficult. His constant rival was the dovish Secretary of State Cyrus Vance, and Carter was torn between them for four years. The Iran hostage crisis broke the presidency. Brzezinski was front and center throughout, pushing harder than Carter was willing to go. He never criticized Carter publicly. Not once, and they remained dear friends for the remainder of their lives.

The Russia obsession can obscure how much else he got right. He was the driving force behind the full normalization of US-China relations in 1979, traveling to Beijing in May 1978 to tell Deng Xiaoping directly that the United States had made up its mind. Shortly after, Deng visited Washington and came to dinner at Brzezinski’s home in Virginia. Brzezinski saw China not as an end in itself but as a strategic counterweight to Moscow. He was equally direct about Iraq. Testifying before the Senate Foreign Relations Committee in February 2007, he called the invasion “a historic, strategic and moral calamity” and warned that staying bogged down would lead to “a head-on conflict with Iran and with much of the world of Islam at large.” He had opposed the invasion from the start. For a man routinely characterized as a hawk, his record on the use of American military force was more nuanced than his critics allowed.

The Paris hotel scene in the early 2000s is one of the more striking passages in the book. Brzezinski was standing in the lobby when an older man in a black leather parka shouted, “Zbiiiiiiig. Zbiiiiiiig!” Gorbachev had been spending most of his time outside Russia, shunned at home, working the lecture circuit. The USSR’s implacable enemy and its final leader, embracing like old friends.

Vindication requires hindsight. Decisions do not have that luxury. Brzezinski’s legacy is not Carter, not Iran, not even Afghanistan. It is the analytical framework: that the Soviet empire was most vulnerable at its edges, that American power on the Eurasian landmass works as a balancing force or not at all, and that a coalition of China, Russia, and Iran united by grievance was the one scenario the West could not afford to allow. The fact that we allowed it anyway suggests his thinking was underweighted.

While not the quality of Sam Tennenhaus’s Buckley, this biography is worth it, and at half the length, it has a succinctness that Brzezinski would respect.

Subscribe now

There’s something uniquely satisfying about a book written by someone who has absolutely no interest in leaving their niche. Not to broaden the audience, not to soften the edges for commercial appeal, but to share an obsession with whoever is willing to follow them down the rabbit hole.

I’m currently reading Mountains of Fire by Clive Oppenheimer, a volcanologist whose charisma is entirely conditional on the topic at hand. Outside his domain, he strikes me as the sort of man who might struggle to hold a room. Give him a volcano, and he becomes electric. There’s a running joke in my family that there isn’t a stranger I can’t bore the moment I start talking about work. I suspect Oppenheimer suffers from the same affliction. Passion, in its purest form, is rarely calibrated for broad appeal.

Tim Gregory belongs firmly in this category. He’s the kind of person you could imagine encountering in a pub in the north of England, quietly unassuming until the conversation turns to nuclear physics, at which point the floodgates open. He even writes about the pub in the book, grounding what could easily be an abstract, technical subject in something recognisably human. If Clive Oppenheimer and Tim Gregory were both at the bar, I’d pull up a stool and stay for the evening. I suspect the conversation would be extraordinary. I don’t expect anyone else to join us.

Going Nuclear is, at its core, a physicist’s attempt to drag nuclear power out of the realm of fear, myth, and ideological distortion and back into the domain of evidence. Gregory’s central argument is simple but forceful: if the objective is rapid decarbonisation at scale, nuclear is not optional; it is essential. What distinguishes the book is not the novelty of that claim, but the methodical way he dismantles the objections that have kept nuclear power politically toxic for decades.

He begins where most anti-nuclear sentiment begins: radiation. Gregory is particularly effective here, not by minimising risk but by reframing it quantitatively. He walks the reader through dose comparisons, background radiation levels, and the measured health impacts of nuclear accidents. The sections on Fukushima and Chernobyl are especially telling, not because they ignore the disasters, but because they separate perception from consequence. The takeaway is uncomfortable: the psychological and political fallout from both events was vastly larger than the physical harm. Fear, it turns out, has been doing more damage than the atom.

From there, he moves to waste, arguably the most persistent rhetorical weapon against nuclear power. His treatment is pragmatic rather than dismissive. Nuclear waste is dangerous, but it is finite, trackable, and already largely contained. Compared to the diffuse, unaccounted waste streams of fossil fuels, the CO₂, the particulates, the slow violence of a warming atmosphere, nuclear waste begins to look less like an existential problem and more like an engineering challenge that societies have chosen not to solve for political reasons.

Where the book becomes most interesting for a financially or strategically minded reader is in its implicit critique of energy system design. Gregory doesn’t attack renewables, but he exposes their structural limitations, intermittency, land use, and the hidden material intensity of scaling them to meet baseload demand. His argument is not that wind and solar are ineffective, but that they are incomplete without firm, low-carbon power to back them up. Nuclear fills that role.

He is less persuasive on economics than on physics, and this is where I’d push back. Gregory acknowledges cost overruns and construction delays but tends to treat them as contingent failures of policy and regulation rather than as systemic features of nuclear deployment in liberal democracies. For an investor or a policymaker, that gap between what nuclear should cost and what it actually costs in the West is not a footnote; it is the entire problem. The book is slightly idealised on this point, and a sharper interrogation of the financing question would have strengthened it considerably.

But that is a narrower criticism than it might sound. What ultimately makes Going Nuclear worth reading is its clarity of framing. Gregory forces the reader into a trade-off that most energy discussions work hard to avoid: if you reject nuclear, you are implicitly accepting either slower decarbonisation or greater reliance on fossil fuels. There is no cost-free alternative. The physics doesn’t negotiate.

There’s a line, often attributed to Tim Ferriss, that you can build a successful business by finding a thousand people willing to pay a thousand dollars a year for something they care deeply about. Books like this exist in that space. They are not written for everyone. They are written for the people who want the weeds. Gregory reads like an evangelist, someone compelled to explain, clarify, and persuade because he genuinely believes the stakes are real.

At just over 300 pages, it’s a quick read that leaves a disproportionate impression. Not because it resolves the debate around nuclear power, it doesn’t, but because it reframes that debate through the lens of someone who has devoted himself to understanding it completely.

More than anything, it makes you want to find him in that pub and continue the conversation. If you see him, ask him about his first book on Meteorites.

Subscribe now

That includes this note.

Price is the only anchor, and while the numbers are clear, what they are telling us is not. Equities are weaker, but the move remains orderly, hardly the behaviour of a market pricing a systemic event. Bond yields have moved, but would central banks respond by tightening in this environment? That depends entirely on jurisdiction. The ECB and the Bank of England, bound more tightly by their mandates, may feel compelled to lean into inflation, while a Federal Reserve led by Kevin Warsh is far less likely to do so. Oil and Natural Gas prices themselves are instructive. They are elevated, but they are not behaving like a market facing a toxic supply crisis. The price of a commodity that cannot be sourced is not marginally higher; it is exponentially higher, and it trades in a disorderly fashion. That is not what we are seeing. Oil remains well below levels that would indicate a true supply dislocation. The market is not pricing closure. It is a pricing disruption, and that distinction matters.

The deterioration in markets over the past week needs to be viewed through that lens. The S&P 500 has made a new low, sentiment has deteriorated further, and the case for near-term stability in risk assets is weaker than it was a week ago. Price is telling you that the path is worse than expected, even if it is not yet telling you that the end state has changed. It is tempting in environments like this to assume that price is confirming the most extreme narratives, that weakness in equities, the move in oil, and the repricing in rates are signalling something structural. More often, they are signalling uncertainty about the path, not certainty about the outcome. What has changed is not the destination; it is the journey.

The deterioration since the last note has three distinct components. First, the geopolitical situation is proving more persistent than initially expected. Iranian asymmetric capabilities, particularly drones and missile systems, have not degraded at the pace the bull case assumed, and disruption in the Strait remains episodic but frequent enough to sustain a higher risk premium.

Second, the US political risk premium has expanded. The Trump administration’s decision-making is harder to read, not just for adversaries but for markets themselves, and that unpredictability is now being priced as a structural feature rather than a temporary distortion. At the same time, domestic political support is beginning to erode in a way that matters for how this conflict evolves. President Trump’s popularity is waning, the MAGA base is showing clear signs of fracture, and influential supporters from the 2024 campaign cycle are stepping away from an administration that ran explicitly on a “no more wars” platform. The attempt to frame engagement with Iran as consistent with an America First doctrine is not resonating and, in many cases, is being rejected outright. That matters because it tightens the political constraint on escalation. This is no longer just about market tolerance; it is about motivating the Republican base to vote in November.

Third, the economic feedback loop is closing faster than expected. Consumer sentiment is weakening, corporate guidance is becoming more cautious, and what was framed as a two- to three-quarter growth slowdown is beginning to show up earlier in the data. Markets are reacting to persistence, not escalation.

What has not changed is more important. A prolonged conflict that drives energy prices materially higher, weakens consumption, and dominates the narrative into the midterm elections is not politically survivable for the incumbent. That arithmetic has not changed, and it places a ceiling on escalation, whether markets choose to recognise it or not. The bar for an off-ramp remains political, not strategic, and the White House will find one.

The interest rate framework also holds. The instinctive reaction to higher oil is to price inflation and therefore tighter policy, but that instinct is increasingly outdated. Higher energy prices function primarily as a tax on consumption, particularly in an economy where services dominate, and labour markets are already softening. The pressure is building as cost burdens intensify, and as that consumption drag becomes visible over the next two to three quarters, the policy response shifts.  Rate cuts become exponentially more likely than hikes. The market is currently pricing inflation risk, but it is not yet pricing the growth deterioration that follows, and the cleanest expression of that mispricing remains in duration.

The bear case is not without merit. Real pressures are building in supply chains, particularly in fertilizer and jet fuel, and those pressures can feed through into food prices, transportation costs, and broader economic friction. There is also a growing narrative that Iran has the upper hand, and that its ability to disrupt shipping and energy flows can be sustained over time.

This is where the analysis begins to overreach.

Iran’s ability to disrupt is real, but its ability to sustain that disruption is conditional. The capabilities that underpin its asymmetric advantage, drones, missiles, and naval harassment, are degradable rather than permanent. That matters, but it is not the whole story.

The Strait of Hormuz is uniquely sensitive, and the threshold for disruption is far lower than most analysis assumes. You do not need sustained closure to create meaningful economic impact. If there is even a sporadic risk of drones or missiles targeting tankers, insurance markets tighten rapidly, premiums spike, and in some cases coverage is withdrawn altogether. Ships will not be crewed if safety cannot be assured. At that point, the constraint is not physical supply, it is financial and logistical. Cargo can exist, but it cannot move.

That dynamic introduces a different kind of fragility. Iran does not need continuous operational control over the Strait to create disruption. It needs to generate a credible enough threat to make counterparties unwilling to take risk. That is a much lower bar, and it is where the bear case has real substance. But disruption is not the same as durable control. Insurance markets reprice, governments intervene, and shipping adapts, albeit imperfectly and at higher cost. The result is not paralysis, but friction. Costs rise, delays increase, and volatility persists, but the system continues to function.

The market is right to take this risk seriously. It is wrong to assume that it translates into sustained control over global energy flows.

That distinction matters for how capital should be positioned, and it brings the discussion back to something more practical. In environments like this, there is an instinct to search for the low, to wait for clarity, to wait for the moment where buying feels justified. That moment does not arrive cleanly. The more important question is not where the market bottoms, but whether your capital structure allows you to act before you know the answer.

If you are managing capital with short-term constraints, the answer is straightforward. You cannot add risk here. Liquidity, mandates, and investor pressure define your strategy, and preserving flexibility matters more than expressing conviction. If you are running a trading-oriented account, the conclusion is similar. Directional positioning requires a view on geopolitical outcomes that cannot be formed with confidence, and that is not an edge.

The only capital that matters in this environment is permanent. The statistical reality of markets has not changed. The largest single-day gains occur in the wake of the largest losses, and waiting for clarity means missing those days. Missing those days compounds into structurally inferior outcomes. Dollar-cost averaging into high-quality compounding assets across a bear market of unknown length is not passive; it is the correct response to uncertainty. The timing of the bottom is unknowable, but the long-term direction of well-chosen assets is not.

The only honest conclusion is that the length of this bear market is unknowable, but the existence of the recovery is not. The current environment feels unstable because it is, and the range of outcomes is wide, but price, the only thing we can trust, is not telling you that the system is breaking. It is telling you that the path is uncertain. Permanent capital that begins allocating into that uncertainty will look back on this period as an opportunity, while capital constrained by liquidity, mandate, or time horizon should wait. The distinction between those two outcomes is not insight or conviction. It is structured. Know which one you are and act accordingly.

Subscribe now

I came to The Technology Trap through its successor. Earlier this year, I read Frey’s How Progress Ends, his sweeping 2025 examination of why nations thrive or stagnate across a thousand years of technological history. It is academic in the best sense: dense with ideas, light on jargon, the kind of book that reorganises how you think about things you assumed you already understood. It made me want to go back to the source.

The Technology Trap, published in 2019, marks the beginning of Frey’s intellectual project in earnest. Its subject is narrower but no less ambitious: how technological change has historically redistributed economic and political power, and what that history tells us about the age of automation we are now living through. It was a Financial Times Book of the Year. It deserved to be.

The book’s analytical spine is a deceptively simple distinction. Frey separates labour-enabling technologies, those that complement workers, raise productivity, and open new avenues of employment, from labour-replacing technologies, those that displace workers faster than new sectors can absorb them. Every major technological transition in history, he argues, can be understood through this lens. The Industrial Revolution was not a uniform tide of progress. It was a prolonged and brutal replacement event, followed eventually by the emergence of new industries and the slow improvement in living standards. The operative word is eventually. Frey’s most sobering data point is what economists call Engels’ Pause: in the period 1780 to 1840, output per worker grew by 46 percent while real wages for the working poor stagnated or fell. Progress happened. Workers paid for it.

The political argument is where Frey is most original. Whether workers lose their jobs to machines is not, he insists, primarily a technological question. It is a political one, determined by who holds power and whose interests shape policy. The Luddites were not irrational. They were skilled craftsmen who correctly understood that mechanisation would destroy their livelihoods. They lost not because they were wrong, but because they lacked the political leverage to slow or redirect the forces arrayed against them. Today’s equivalent, the despairing middle class turning toward populism and political fragmentation, is, in Frey’s reading, the predictable consequence of allowing disruption to outrun institutional response. His “technology trap” is the danger that resistance to short-term pain forecloses long-run gain. But in 2025, a second trap has become visible that Frey did not fully anticipate: the one where societies embrace new technology without any redistributive infrastructure in place to catch those left behind.

This is where the book’s vintage begins to matter. The Technology Trap was written in 2019, before ChatGPT, before the boardroom consensus that large language models would reshape white-collar work at scale, before the question of cognitive displacement became impossible to dismiss. In that context, Frey was ahead of the curve, warning about political backlash, stagnating middle incomes, and the inadequacy of historical optimism, at a moment when most economists were still broadly sanguine. That prescience deserves acknowledgment.

But the book rests on an assumption it never fully examines: that Schumpeterian creative destruction will, in time, do what it has always done. New sectors will emerge. That displaced workers, given sufficient time and retraining, will find their way to them. Every prior wave of automation destroyed old categories of work and created new ones. Factory hands replaced handloom weavers, service workers replaced factory hands, and knowledge workers replaced clerical labour. The historical record seemed to validate the pattern.

Reading The Technology Trap in 2026, that pattern is the thing I most want to interrogate. AI is categorically different from prior automation waves, not in degree but in domain. Every previous technology attacked a bounded category of work. AI attacks cognition itself across sectors simultaneously: the junior analyst, the paralegal, the radiologist, the copywriter. There is no cognitive tier left to retreat to and unlike the railway boom or the electrification of industry, we cannot yet point to the new sector being created to absorb those displaced. Green energy, care work, and the creative industries are real but insufficient in scale and not themselves immune to AI encroachment. The absorptive mechanism that underwrites Frey’s implicit optimism is precisely the thing now most in question.

None of this diminishes what Frey built. The Technology Trap remains the most rigorous historical account of how technology reshapes labour markets and political power. Read it for that. But read it also as a transitional text, the best framework the pre-AGI intellectual world produced, asking all the right questions, arriving at answers that the speed of events is now forcing us to revise. If Schumpeterian absorption cannot be relied upon, the question Frey leaves open becomes urgent: who captures the productivity gains of AI, and by what mechanism do they flow back to those who bear the cost of disruption? That is not a technological question. It is, as Frey would agree, entirely a political one.

Subscribe now

One of my pet peeves of the investment research/pundit industry is the flat refusal to indicate a time frame when discussing being bearish. The default trajectory of equities is up. That is undeniable. In the past 75 years, rolling one-year total returns have been positive for US equities 73% of the time. It is 90% of rolling 5-year total returns, 97% of rolling 10-year total returns, and 100% of rolling 20-year total returns. So being bearish requires nuance.

It is okay to be bearish, but a negative stance without a time frame for when it plays out is not very useful as you face an inevitable long-term propensity to own companies. Being bearish over the next week, month, or year can be completely defensible, whether based on valuations, earnings expectations, or concerns about the economic trajectory. A blanket bearish proclamation without timing is like buying an option without knowing the expiry.

I raise this because the Iran War has generated an enormous volume of exactly this kind of undated, unqualified bearishness. The instinct to reach for the macro panic button is understandable. A US-Israeli strike that killed the Supreme Leader, retaliatory missile exchanges across the region, and the spectre of disruption to the Strait of Hormuz are not trivial events. But my job is to identify what has changed in the cash flow outlook, over what time horizon, and whether the market reaction is pricing something durable or something transient. In most geopolitical episodes, it is the latter. Iran is more complicated than most, but the same discipline applies.

This piece is a little easier to write on Monday, following President Trump’s announcement of productive talks with the new leadership of the Iranian regime. I, like many, had major concerns about the US military offramp: the scenarios that would allow the US to claim victory in order for the administration to appease growing disenchantment at home with a voting base that is against the war and certainly against the economic consequences. This looks like a continuation of Trump Always Chickens Out (TACO), the idea that adverse market reactions are the only true guardrails for the President’s policy agenda. In fact, it can be argued that this is the most important analytical framework for understanding the ceiling on this conflict, and at a minimum, the announcement earlier today and the resultant collapse in energy prices appear to be a turning point.

What matters for capital markets is that the pattern is consistent and, crucially, it is markets that appear to have again forced this latest pivot. The Greenland episode is the clearest illustration. President Trump retreated from what was widely viewed as an untenable assertion of US sovereignty, not because of diplomatic pushback, but because markets made the financial consequences apparent. As I noted in a February report, President Trump himself alluded to adverse market reactions to his Greenland ambitions during his Davos address. The same dynamic applied to tariff announcements that were walked back within the news cycle once bond markets began to price the consequences. US equity, bond markets, and now energy prices remain the ultimate arbiters of the life expectancy of any Trump policy, and the White House has demonstrated a clear and consistent sensitivity to that feedback mechanism.

Applied to the Iran War, TACO operates as a structural ceiling on escalation risk. The “stated objective”: the degradation of Iran’s missile and nuclear architecture, has been substantially achieved. What the administration does not want, and will actively work to avoid, is a ground campaign in Iran, a multi-week closure of the Straits of Hormuz that drives oil above $130, or a conflict that becomes the defining domestic political narrative heading into the midterms.

The electoral arithmetic is unambiguous: Historically, the president’s party loses seats in midterms, and the odds heavily favoured Republicans losing the House of Representatives. A protracted Middle East war with $5 gasoline is not the environment in which that arithmetic improves. Israel’s objectives and Washington’s partially overlap here, but they are far from synchronized. Netanyahu’s aim appears to be the obliteration of the regime with whatever tools are necessary. The Trump administration’s goals are less defined, but US engagement is not infinite.

That divergence matters for how this evolves. TACO does not prevent the near-term volatility, but it does cap the structural damage and should remove the most catastrophic tail scenarios from the realistic probability distribution, despite the complexity of the situation.

The Economic Consequences are Complicated

Twenty years ago, the answer to the question of oil’s impact on inflation was not in serious dispute. Oil up, CPI up. The relationship was direct, well-documented, and the Federal Reserve understood how to respond. That relationship has been structurally weakened over the past two decades, and the current debate underprices the extent to which it has changed. The US economy is significantly more services-oriented than it was during the inflation era of the 1970s or even the early 2000s. Energy’s share of the CPI basket has compressed. The pass-through from crude oil prices to core services inflation, wages, rents, etc., is indirect and slow, and the Fed’s own trimmed-mean measures consistently show that oil-driven headline moves tend to revert without durably embedding in core if labour markets are not simultaneously tightening.

Higher oil prices still act as a tax on US consumers, squeeze oil-intensive sectors, and lift headline inflation, but the national effect is more mixed than it was before the shale revolution because a meaningful share of the income transfer stays inside the domestic economy through producers, wages, profits, and capex.

The sector arithmetic still matters. Energy is only a small slice of the S&P 500, roughly mid-single digits by market cap, so the direct beneficiaries of a sustained oil spike remain far smaller than the broad swath of the index that faces a margin headwind. A $20/barrel increase in Brent should translate into approximately 48 cents per gallon at the pump, according to the American Energy Information Administration, and that is enough to meaningfully erode household disposable income, particularly for the lower end of the income distribution.

That last point is the most important. The burden is regressive. Lower-income households devote a larger share of spending to gasoline and household energy, which means the hit lands where consumption is most price-sensitive and savings buffers are thinnest. So even if higher oil prices are no longer the clean macro negative they once were for the US in aggregate, they are still a meaningful drag on consumption, politics, and market sentiment.”

The current environment is best described as a tax-versus-inflation debate rather than a straightforward stagflationary spiral. An oil price spike in a softening labour market slows growth without durably lifting core inflation. It presents as stagflationary on the surface, growth lower, headline CPI higher, but the underlying dynamic is disinflationary in the core, because the growth drag eventually outweighs the energy pass-through. In theory, a Federal Reserve that understands this dynamic should be tolerant of headline CPI spikes when core remains anchored, and the labour market is weakening. Whether this Fed, with a new chair being confirmed in a politically charged environment, will have the institutional confidence to act on that understanding is an open question, and one with significant implications for how global interest rates trade over the remainder of 2026.

Curve Repricing

US front-end yields have moved, and the market is attempting to price two contradictory things simultaneously: an inflation risk premium that argues for higher rates and a safe-haven bid that argues for lower ones. That internal contradiction is unlikely to persist in one direction for long. My view is that near-term repricing in the US is justified on the headline inflation story but structurally incorrect once the growth drag feeds through over the next two to three quarters. The 2s10s curve should steepen from here, not because the long end sells off aggressively, but because the front end rallies as the consumption squeeze becomes impossible to ignore in the data.

The more interesting comparison is between the US and oil-importing emerging markets. Countries like India, Turkey, and the Philippines face the identical consumption tax shock but with a fraction of the fiscal and monetary flexibility to absorb it. When oil rises sharply in these economies, current account deficits widen, currencies come under pressure, and central banks face the classic impossible trinity: defend the currency and accept a growth hit, cut rates to support growth and watch the currency weaken further, or allow inflation to run. There is no clean option. The EM front-end repricing is substantially more warranted than the US equivalent, and it has structural duration rather than being a short-term noise event.

Europe occupies an intermediate position. It is heavily dependent on energy imports; Germany is already in an industrial recession, and a sustained oil shock would hit GDP before it reaches core inflation. The ECB has less room to cut than its dove

Subscribe now

s will admit, and more political pressure to do so than the market currently prices in. I expect the European front-end repricing to ultimately ease as growth disappoints, making it a near-term event rather than a durable shift in the rate structure.

The UK sits in an uncomfortable position that is distinct from the continental European story. The Bank of England is navigating a combination of sticky services inflation and a labour market that is softening faster than the MPC’s models anticipated. An oil shock complicates the calculus in both directions simultaneously: it pushes headline CPI higher at exactly the moment the growth outlook is deteriorating, making the rate path ambiguous in a way that is not true for the ECB. Sterling’s sensitivity to global risk appetite adds a currency dimension that the eurozone does not face in the same way. The UK front-end repricing is justified in the near term, but I would expect it to reverse as growth data deteriorates through the second half of 2026, broadly following the European trajectory.

Japan is a different case entirely and is worth separating from the broader developed-market discussion. The Bank of Japan has spent three decades trying to generate durable inflation and is now, for the first time in a generation, in a tightening posture. An oil shock creates genuine tension in Tokyo: higher energy import costs are unambiguously negative for Japan’s terms of trade and for real household income, but the headline inflation impulse feeds into a BoJ framework that has been using any available inflation as justification for normalisation. Governor Ueda is unlikely to treat an oil-driven spike in CPI as a reason to accelerate tightening. The BoJ understands the distinction between imported and domestically generated inflation better than most, but the yen’s reaction function to oil is negative, which creates its own feedback loop for Japanese asset prices in the near term.

The repricing risk becomes more nuanced in commodity-sensitive emerging markets that sit at the intersection of strong terms of trade and tight global liquidity. Brazil is a good example.

At first glance, higher oil prices should be supportive. Petrobras is a meaningful contributor to fiscal revenues, and Brazil’s broader commodity complex benefits from a firmer global pricing environment. But the transmission is not clean. Domestic fuel pricing remains politically mediated, and the currency has been under intermittent pressure in a strong-dollar environment.

That leaves policymakers managing a familiar tension, albeit from a position of relative strength. Real rates are already restrictive, and the front end reflects that. In a scenario where oil strength reinforces a firm dollar and tighter global financial conditions, the trade-off becomes one of calibration rather than crisis: how much to lean against currency weakness versus how quickly to support domestic growth.

This is not a fragile equilibrium, but it is an uncomfortable one. Brazil is better positioned than most to absorb a commodity shock, yet the combination of tight domestic policy and externally driven currency pressure means the adjustment is unlikely to be smooth. The curve reflects a degree of caution but may still underestimate how persistent that tension could prove to be if dollar strength endures.

South Africa is the cleaner expression of this risk. The rand is one of the most liquid high-beta currencies in emerging markets, so it tends to respond quickly to global risk aversion. South Africa does benefit from higher gold and platinum-group metal prices, but persistent electricity and logistics constraints blunt the transmission of stronger commodity prices to stronger domestic activity. That leaves the SARB in a familiar position: with inflation credibility hard-won and the new 3% target still anchored, it is likely to be cautious about easing into renewed currency weakness, even if domestic growth remains soft. If EM stress persists into the second half of 2026, South Africa looks more exposed than Brazil to that tension between external vulnerability and domestic stagnation.

The Cost of Missing the Recovery

Days like Monday do more damage to longer-term performance than the underlying sell-off does because it is these sorts of sessions where underexposure becomes a compounding error that is rarely recovered from.

JP Morgan Asset Management has done the work on this, and the conclusion is one that every investor tempted to go to cash during a geopolitical shock should keep close to hand. Seven of the ten best single-day returns in the S&P 500 over the past twenty years occurred within two weeks of the ten worst days. Not in the weeks after clarity arrived. Not after the ceasefire, the peace deal, or the reassuring press conference. Within a fortnight of the most frightening moments in markets.

The compounding consequences of missing those recovery days are severe. An investor who put $10,000 into the S&P 500 in January 2002 and held through every crisis, every war, and every collapse over the following twenty years ended with more than $60,000 by January 2022, a 9.52% annualised return. An investor who held the identical position but missed only the ten best days over those twenty years, ten days out of approximately 5,000 trading sessions, or 0.2% of available time, ended with less than $30,000. Missing 0.2% of the period more than halved the terminal wealth.

The mechanism behind this pattern is directly relevant to the current environment. Fear-driven selling creates dislocations, which are followed by large, rapid recoveries before the news flow allows investors to feel comfortable re-entering. By the time Iran looks safer, by the time oil flows again through the Straits of Hormuz, by the time the diplomatic framework is in place, the majority of the recovery is already priced in. This is not an argument for ignoring risk. It is an argument for understanding that the risk of being out of the market during a recovery is at least as consequential as the risk of being in the market during a drawdown, and it is far less visible and psychologically present when the decision is being made.

Seeking an off-ramp

The Iran War is a serious geopolitical shock, but it is not yet an economic catastrophe, and that distinction matters enormously for positioning.

The United States has absorbed oil shocks before. It has absorbed wars before. What it has consistently avoided is allowing a foreign policy commitment to run so far ahead of domestic political reality that it becomes economically or electorally destabilising. President Trump understands that instinctively. Whatever one thinks of the decision to support the Israeli strike, the probability that the administration allows this conflict to evolve into a prolonged, growth-damaging event heading into the midterms remains low. TACO is not a weakness. It is a constraint, and it applies here as much as it did with Greenland and tariffs.

The supply chain disruption risk is real, but it is being overstated. Comparisons to COVID are misplaced. This is not a simultaneous collapse in global demand and supply with no policy framework. It is a regional shock to energy infrastructure and shipping routes that, while disruptive, is navigable. Businesses can adapt, inventories can adjust, and alternative routing exists. The scale is different, and so is the economic consequence.

The AI narrative has not changed in any meaningful way. The capex cycle that is restructuring corporate cost bases, the productivity gains beginning to emerge in margins, and the multi-year earnings trajectory that underpins the constructive view on equities remain intact. Geopolitical risk can interrupt positioning, but it does not alter that underlying trend.

The strategy follows from that framework. Use weakness to add exposure, particularly in areas where long-duration earnings are being sold for short-duration fear. That does not mean ignoring near-term volatility or the more persistent headwinds facing parts of emerging markets. It does mean recognising that the default trajectory of equities remains higher, and that waiting for clarity in an environment like this is rarely rewarded.

The investors who will look back at this period and regret their positioning will not be those who selectively added risk into dislocation. They will be those who stepped aside entirely, waiting for a resolution that markets had already begun to price.