The Sincerest Flattery

The idea

This inquiry began with a war. On October 6, 2026, The Wall Street Journal published Alistair MacDonald and Nikita Nikolaienko's report, "The Silent Electronic Battleground Shaping the Future of Warfare," on the contest of jamming and counter-jamming in Ukraine. On that front the best tools are the first to be answered. Britain's Royal United Services Institute found that a drone needs new software, sensors and radios every six to twelve weeks as the enemy learns to defeat them. One Ukrainian officer has described needing a single jamming module early in 2024, four or five by that summer and seven by fall, before drones guided by fiber-optic cable made jamming largely beside the point. The director of the Pentagon's Defense Innovation Unit told Congress in 2025 that its drones must be updated "within weeks or even days, or risk those capabilities being obsolete before even being produced, much less deployed."

The article left me with a question, which I took to Vista's Desk that same afternoon: can the best technology look worse? On the battlefield, the most effective jammer is the one the enemy works hardest to defeat, so it is often the first to stop working. Could the same thing happen in business? Picture two companies, each with a product ahead of the competition. One is a little ahead. The other is far ahead, with a breakthrough. The breakthrough is worth far more to any rival who can copy it, so rivals pour money and talent into catching up with it and leave the modest product alone. A few years later the breakthrough's lead may have shrunk to almost nothing while the modest product's lead is intact. Anyone who judged the two by how far ahead each still was would pick the wrong one as the better technology, even though the breakthrough made far more money along the way. The Desk challenged the idea over several rounds, and a short manuscript followed, drafted with the help of another AI assistant, with a small model of the chase at its center. This inquiry takes that model out of the lab and into the markets.

Highland Park, 1914

The best technology of the twentieth century was given away.

"Along about April 1, 1913, we first tried the experiment of an assembly line," Henry Ford wrote. "We tried it on assembling the flywheel magneto." One man, doing the whole job, had made a magneto in about twenty minutes. Spread along a moving line in twenty-nine operations, it took thirteen minutes and ten seconds; raised eight inches, seven; at the right speed, five. Then Ford moved the whole car. In September 1913 the best stationary assembly of a Model T chassis took twelve hours and twenty-eight minutes of labor. On April 30, 1914, the Highland Park plant in Michigan assembled 1,219 chassis in one eight-hour day, at an hour and thirty-three minutes each, an eightfold saving, and a finished chassis came off its lines about every twenty-four seconds. The engineers who recorded it, Horace Arnold and Fay Faurote, summed it up in a single line: "The great secret of Ford's labor-cost reductions is the moving assembly."

It was no secret at all. Arnold opened his account, serialized in The Engineering Magazine from April 1914 and published as a book in 1915, with the most remarkable fact of the whole enterprise: "the Ford company is willing to have any part of its commercial, managerial or mechanical practice given full and unrestricted publicity in print." So many visitors walked the plant that its aisles had to be marked. Ford said it plainly a decade later: "For we have no trade secrets. If we are doing anything which another manufacturer may find use for, then we want that manufacturer to have the benefit of what knowledge we possess. That we take as our duty." He was not naive about it. He bought one of every new car his rivals made, drove it, took it apart and studied how everything was made.

In January 1914 Ford announced the five-dollar day, more than double the old rate, and the next morning, United Press reported, "Ten thousand anxious, determined men, some ragged and unkempt, others seemingly prosperous," fought for places in the line outside the plant, and within days, in icy weather, the crowds were being turned back with fire hoses. The price of a Model T touring car fell from $950 in 1909-10 to $290 in 1925, and Ford built fifteen million of them. In 1924, by my arithmetic from the production records, Ford made about 55% of all the passenger cars built in the United States.

The industry did what Ford invited it to do. Moving lines were running at a number of other automakers by the middle of 1916. The whole system took longer to copy, about a decade, and it arrived at General Motors with a defector: William Knudsen, who had run production for Ford, went to Chevrolet and promised to match Ford, as he said in his Danish accent, "vun for vun." Alfred Sloan's General Motors then did something Ford would not, offering "A Car for Every Purse and Purpose" and a new model every year, while Ford insisted, "We never make an improvement that renders any previous model obsolete." Ford's share fell to 44% in 1925 and 36% in 1926. In 1927, the year the Model T was retired, Chevrolet outsold Ford.

So by any late measure Ford's lead looked worse than a lesser invention's might have. Its share was falling and its rivals were catching up. By any measure of what the lead had earned, nothing in the industry came close. Ford's technology was the best, it was the most worth copying, and it was copied hardest. (Ford built back, and was on top again in 1929. The point is the shape of the lead, not the fate of the firm.)

The Wright brothers took the opposite road and arrived at the same place. They kept their methods close and fought for their patents in court. Sir Walter Raleigh, who wrote the official history of Britain's air war, observed in 1922 that after Wilbur Wright flew in France in 1908, "the progress of flying was rapid and immense. A great industry came into being, and, after a short time, ceased to pay any tribute whatever to the inventors." Between 1908 and 1913 the American government spent $435,000 on aviation and Germany $28 million, and the American fighter pilots of World War I flew European aircraft. Openness and secrecy, the two ends of the choice, lost the lead the same way.

Charles Caleb Colton summed it up in 1820: "IMITATION is the sincerest of flattery." Colton, a clergyman whom the Dictionary of National Biography called "more famous as a sportsman, and especially as a skilful fisherman, than as a divine," was himself charged with borrowing from Bacon.

The Red Queen

This reminds me of the Red Queen. In Through the Looking-Glass, Lewis Carroll's Alice runs as fast as she can, hand in hand with the Queen, and finds herself under the same tree. "Now, here, you see," the Queen explains, "it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!" In 1973 the biologist Leigh Van Valen gave the name to a law of evolution: because every species' gain worsens the world of its rivals, each must keep adapting merely to survive, and none wins for long. Several journals turned the paper down, Nature among them, so Van Valen founded a journal of his own and published it on page one of volume one, thanking the National Science Foundation "for regularly rejecting my (honest) grant applications." Students of business took up the idea two decades later. William Barnett and Morten Hansen found that a firm's learning provokes its rivals' learning, which comes back as pressure on the firm; a later study of more than 4,700 competitive moves found that each one lifted the mover's results and also hurried its rivals' replies, which then cut them.

The Red Queen says that everyone has to keep running. It does not say who will be chased the hardest. That is the question the manuscript's model answers.

The catch-up arch

The manuscript's model, which we call the catch-up arch, fits in a few sentences. A company's product is ahead of its rivals': it works better, costs less, or does something they cannot. Call the size of that advantage the company's lead. A rival decides how hard to try to catch up. Trying harder costs more. The reward for catching up is bigger when the lead is bigger, because the leader's customers and profits are what the rival stands to win. So the rival's effort, and the share of the lead it takes away, grow with the size of the lead.

Let q be the size of the lead, from 0 (no advantage at all) to 1 (the largest advantage there could be). Let g be the prize to a rival for each unit of lead it captures, and c the cost of catching up. The rival chooses a, the chance that it catches up, to make its gain as large as it can be:

g × q × a − c × a² / 2

The first term is the prize it expects; the second is the cost, which rises steeply the harder it tries. The best choice is a = k x q, with k = g / c, the prize for catching up divided by its cost, and never more than certainty: a = min(1, k x q). The lead that survives is the part not caught:

R(q) = q × (1 − a) = q × (1 − k × q)

And the value the leader earns along the way, before the rival arrives and after, is

V(q) = b × q + d × R(q)

where b weighs the early years and d the later ones. Try it with k = 1 and both weights at 1:

Compare a lead of 0.4 with a lead of 0.8. The bigger lead earns 50% more, 0.96 against 0.64, and at the end keeps only two thirds as much of its advantage, 0.16 against 0.24. Anyone who measured the two at the end, by the lead that remained, would rank them backwards. That is the whole paradox, and it is why we call the model an arch: the lead that survives rises with the size of the lead, peaks at q = 1 / (2k), where it equals 1 / (4k), and falls to nothing at q = 1 / k, where the chase becomes a certainty.

The size of the prize moves the arch. When k is a half, because the prize is small or copying is costly, the arch peaks at the very top, and the best technology keeps the most of its lead. When k is 2, the peak falls to a quality of 0.25 and every lead above 0.5 is caught completely. When k is 5, as it might be in a market as large as the one for artificial intelligence, only leads below 0.2 survive at all. The same lead, in a bigger market, is chased harder.

War is the limiting case. It is where this inquiry began. On the front in Ukraine the prize for answering a new weapon is survival itself, and copying a radio frequency costs almost nothing, so k is about as large as it can be and only the smallest leads last. That is why the advantage has passed from the best weapon to the fastest cycle. Both armies now redesign their drones every few weeks, and the Pentagon's own testimony gives the scale: "In 2025, DoD plans to buy about 4,000 drones. In Ukraine, more than 4,000 drones are produced and consumed per day." When even a modest lead is caught within weeks, the only lasting edge is the speed of the next change.

One condition matters more than any other, and the record insists on it. The arch appears when the cost of copying stays roughly the same as the lead grows. If copying gets more expensive as fast as the lead grows, the arch disappears. Let the cost of catching up rise with the size of the lead raised to the power gamma, and hold the prize-to-cost ratio at q = 1 the same, 0.8, in every case:

When copying costs about the same whatever the prize (gamma near zero), the best is chased hardest. When copying costs rise as fast as the lead (gamma of one or more), the best keeps the most. So the investor's question about any lead is a pair of questions: how much is it worth to the best-funded rival, and does copying it get more expensive as it grows?

And here is the measurement trap. The share of a lead that survives is 1 - k x q. A long record of stable share and high returns, the usual evidence of a moat, measures k x q, the prize-to-cost ratio times the size of the lead, not the size of the lead alone. It cannot tell a lead that rivals found too hard to copy from a lead they found not worth copying, and it can rank the most valuable leads below the middling ones.

What the record says

Economists have measured the chase for half a century, and most of what they found fits the arch. In 1981 Edwin Mansfield and two colleagues studied 48 new products and found that copying one cost its imitators, on average, about 65% of what it had cost the innovator and took about 70% of the time; about 60% of the patented successes were imitated within four years. Four years later Mansfield reported that the details of a new product or process usually leak to rivals within about a year. In 1987 Richard Levin and three colleagues surveyed 650 research managers in 130 lines of business: they rated lead time, the learning curve and sales effort as better protection than patents, and said that a typical unpatented product was copied within a year in two industries out of three.

The prize shows up in the chase wherever anyone has looked for it. Henry Grabowski and Margaret Kyle followed 251 drugs as they first met generic competition between 1995 and 2005: drugs selling under $100 million a year had kept the market to themselves for about fifteen or sixteen years, those selling $250 million to $500 million for about ten, and the biggest drugs drew about four times as many generic entrants in their first year. R. David McLean and Jeffrey Pontiff studied 97 published predictors of stock returns and found their returns 58% lower after publication; the predictors that had paid the most lost the most. Every investor who has watched a good strategy crowd out knows that paper. David Teece described the pioneer's lament in 1986, in a paper that has shaped how a generation thinks about who profits from an invention. EMI introduced the CT scanner in the United States in 1973, lost the lead within six years and was out of the business by the eighth; RC Cola sold the first cola in a can and the first diet cola, and Coke and Pepsi followed almost at once. I have known David for more than thirty years. When I founded Round Table Group in 1994, he was one of the first scholars to sign up and one of its earliest and strongest advocates. We have been friends and business associates ever since, and a few years ago we founded the Silicon Valley Innovation Institute together.

I know the pioneer's lament from the inside. In 1994 I founded what was arguably the first expert network. As I wrote in my book, The Scholars' Treasure (chapter 7, page 95), around 2000, while I was distracted by the dot-com boom, one competitor was "taking our model and applying it to financial services." The firms that followed built what David calls complementary assets: the sales forces, the client relationships and the scale. They kept most of the value of the idea. David had explained why in 1986: when an idea is easy to copy, the profit goes to whoever owns what the copier cannot copy.

William Nordhaus later estimated that innovators keep about 2.2% of the social value of what they invent.

Not all of it fits. Levin's survey also found that major patented products were rarely copied within a year and that only three to five firms could copy them at all. Dennis Mueller found that the most profitable of 600 large American manufacturers in 1950 faded the most slowly over the next two decades. Those are the cases where copying gets more expensive as the lead grows: the complex systems, the large firms with assets a rival cannot buy. The model does not dispute them. It says they are the gamma of one or more, and that they are not the whole market.

How investors judge a lasting advantage

Vista's Decision Library carries the investor's standard tool for this, under the title "How long a high return lasts." Its base rate comes from Eugene Fama and Kenneth French, who reported in 2000 that a company's profitability moves back toward the average by about 38% of the gap each year, so that half of any gap is gone in under a year and a half. That was my last year at the University of Chicago, where I took every finance and investments course on offer while Fama and his colleagues were teaching there. Michael Mauboussin and Dan Callahan, studying American companies from 1970 to 2024, put the yearly fade between 10% and 30% by sector, about 21% on average. Morningstar's analysts give a company a wide moat when they expect excess returns to last at least ten years, and probably twenty, and set the fade in their valuations accordingly. Bruce Greenwald's two tests of an advantage are stable market shares and high returns.

Every one of these reads persistence as the evidence. The Library entry carries its own warning, that the average fade is not a law for every company. The model says which way each kind of company departs from it. Where the prize is enormous and copying costs about the same whatever the prize, the fade will run faster than any average, and a long record of high returns says more about the size of the prize than the quality of the lead. Where the prize is small, a lead can outlast every average, because no one thinks it worth the chase. The average is too slow for the leads that matter most to the market and too fast for leads it barely notices. None of this settles the old argument over whether pioneers or fast followers win: Ford, the pioneer of mass production, held his lead for a decade, while Lilly, a follower, took the lead in weight-loss drugs. What the arch adds is narrower and, as far as our search of the literature found, new. A lead that lasts is evidence about the size of the prize as much as about the quality of the product, and a moat measured by persistence can rank the most valuable leads below the middling ones.

The price of thought

Artificial intelligence is the cleanest test of the arch yet devised, because copying a model can cost almost nothing next to the prize. OpenAI described o1, the first of its models trained to reason step by step, on September 12, 2024, and released it to developers in December at $15 for every million tokens, the word fragments a model reads, and $60 for every million it writes. On January 20, 2025, 130 days after the announcement, the Chinese lab DeepSeek released R1, which it described as on par with o1:

R1 charged $0.55 and $2.19, about a twenty-seventh of o1's price. Its reasoning training, DeepSeek later reported in Nature, cost about $294,000, on top of a base model whose final training run cost about $5.6 million.

The chase has been measured since. Epoch AI, a research group that tracks the field, reported on September 22, 2026 that the cost of reaching a given level of performance has fallen about 47% a quarter since 2023, about thirteen times a year, and that it falls fastest when the level is new: about 75 times a year in the first year after a level becomes the best there is, and about 4.7 times a year two years later. Researchers at MIT found prices falling about 32 times a year for the best tier of models and about 1.7 times for the weakest. Open models now trail the best closed ones by about four months. In AI the best technology looks worse fastest.

This reminds me of King Solomon, to whom tradition credits the book of Ecclesiastes: "What has been will be again, and what has been done will be done again; there is nothing new under the sun. Is there a case where one can say, 'Look, this is new'? It has already existed in the ages before us." (Ecclesiastes 1:9 to 10, Berean Standard Bible.) Solomon was speaking of the vanity of all things. In artificial intelligence he could have been describing a business model. What is new today is common in about four months and cheap within a year.

Who keeps the gains, then? David Teece's answer was the owners of complementary assets, the things a copier cannot copy: customers, distribution, and capacity. The latest filings of the companies that sell AI capacity are suggestive. From the second quarter of 2025 to the second of 2026, the operating margin of Google Cloud rose from 21% to 36% and that of Amazon Web Services from 33% to 39%; both offer AI chips of their own design, Google its TPUs and Amazon its Trainium. Over the same span Meta's operating margin fell from 43% to 31% as it spent on capacity, Oracle's gross margin from 67% to 60% and CoreWeave's from 74% to 66%, and Microsoft's cloud gross margin had slipped from 72% in fiscal 2023 to 66% in fiscal 2026. The filings do not say that the chips are the reason, and many things move a margin. But the pattern is the one Teece would have predicted.

The first-mover's fall

The arch has already run its course in one great market. Novo Nordisk's semaglutide, sold as Ozempic for diabetes and Wegovy for weight loss, was the best technology of its kind when Wegovy was approved on June 4, 2021. The prize was the largest in pharmaceutical history, and it drew the hardest chase. Eli Lilly's tirzepatide, approved as Mounjaro in May 2022 and as Zepbound in November 2023, sold about a seventh as much as Novo's semaglutide in the first quarter of 2023, pulled even in the second quarter of 2025, and in the third quarter of 2025, about four years and four months after Wegovy's approval, passed it clearly, $10.1 billion against $8.9 billion. By the second quarter of 2026 Lilly's franchise was selling 1.61 times as much as Novo's, and Lilly held about 61% of American prescriptions for the class against Novo's 39%.

Novo's market value, which had passed $600 billion in the summer of 2024, is now about $165 billion. The company issued an unscheduled profit warning and replaced its CEO in July 2025, told investors in February 2026 to expect sales to fall that year, and has cut the American list prices of Wegovy and Ozempic by about half and a third from January 2027. Measured late, by share or by market value, Novo's lead looks worse than almost any in the industry. Measured by what it earned, few in history have matched it.

The race has not stopped at the new leader. Lilly's share of prescriptions has held at about 60% to 61% for three quarters, as Novo has launched a pill and cut its prices. If the model is right, the most valuable lead in medicine will now be the hardest chased, whoever holds it.

The most valuable lead in technology

How NVIDIA came to hold its lead is itself a story of the arch. Jensen Huang, Chris Malachowsky and Curtis Priem planned the company in 1993 in a booth at a Denny's on Berryessa Road in San Jose, where the restaurant chain unveiled a plaque thirty years later. Their first chip, the NV1, bet on the wrong standard. The industry chose Microsoft's and Silicon Graphics' instead, sales of the NV1 stopped, and NVIDIA lost money in every quarter until the end of 1997. Huang has said that Sega, the customer whose console chip NVIDIA had to abandon, was generous enough to give the company about six months to live. The RIVA 128 saved it. In 1999, the year NVIDIA went public at $12 a share, it introduced the GeForce 256 as the world's first GPU.

Then came the bet that built the lead. In November 2006 NVIDIA introduced CUDA, which let programmers use its graphics chips for general computing. The company called it the industry's first C-compiler development environment for the GPU and pitched it for product design, data analysis and physics, not for artificial intelligence. It was expensive. Huang has said CUDA raised NVIDIA's costs by half when its gross margin was about 35%, and that shareholders wanted the company to focus on profits instead; the gross margin fell from 45.6% in fiscal 2008 to 34.3% in fiscal 2009, with the recession playing its part. In the model's terms, the lead could be built because the prize still looked small. For six years, computing on graphics chips was a niche for scientists and engineers, and the chase that follows a large prize had not begun.

The prize appeared in 2012. Alex Krizhevsky, working with Ilya Sutskever and Geoffrey Hinton, trained a neural network called AlexNet on two NVIDIA graphics cards in his bedroom at his parents' house, using his own CUDA code, and won that year's ImageNet image-recognition contest with an error rate of 15.3%, against 26.2% for the runner-up. NVIDIA's annual report now calls it the big bang of AI. In 2016 Huang carried the first DGX-1, a $129,000 deep-learning computer, to OpenAI's offices in San Francisco himself. In 2019 NVIDIA agreed to buy Mellanox, whose networking ties enormous numbers of chips into a single machine, for about $6.9 billion. Its data center business passed gaming for good in the spring of 2022, six months before ChatGPT. Today 7.5 million developers use CUDA and NVIDIA's other software tools. The lead was built over a decade, in a quiet market, by a company the chase had not yet noticed. It was an NVE before it was an NVIDIA.

No lead in the history of business has been worth more than NVIDIA's. Its data center revenue was $4.3 billion in the quarter that ended in April 2023 and $89.0 billion in the quarter that ended in July 2026. Its gross margin has held at about 75%. On October 6, 2026 its shares closed at $239.24, a record, and the company was worth about $5.8 trillion.

By the model, a lead this valuable should draw the hardest chase in technology. It has. Broadcom, which designs custom accelerators for the largest cloud companies, earned $16.7 billion from AI chips in the quarter that ended in August 2026, up 221% from a year before, and expects $21.7 billion in the next. Through the first quarter of 2026 its AI revenue ran at 11% to 14% of NVIDIA's data center revenue; in the second quarter it reached 18.8%. IDC, which tracks server sales, reports that servers built around custom chips rather than graphics processors took about 11% of spending on accelerated servers in the second quarter of 2025 and about 24% a year later. Google began selling its TPU systems to outside customers in the second quarter of 2026, and Anthropic has agreed to take about 3.5 gigawatts of TPU capacity through Broadcom from 2027. Amazon reports more than $225 billion of commitments for its Trainium chips. Meta is building its own accelerators with Broadcom; Microsoft says its Maia 200 delivers 30% better performance per dollar than the latest hardware in its fleet; OpenAI has ordered ten gigawatts of accelerators of its own design from Broadcom. AMD's data center revenue in the quarter that ended in June 2026 was more than double its level a year earlier. NVIDIA has answered in kind, licensing the inference technology of the startup Groq at the end of 2025.

Why, then, is NVIDIA still winning? For three reasons, and none of them contradicts the arch. First, the chase started late. NVIDIA built its lead for a decade while no one was running after it, and the custom chips now taking share were ordered at scale only in 2025 and 2026; the record says copying costs most of what inventing did and takes most of the time. Second, the market is growing so fast that the leader and the chasers can all grow at once, so the chase shows up in shares before it shows up in sales. NVIDIA's share of the AI chip revenue of NVIDIA, AMD and Broadcom together slipped from 83% to 79% in the past year, while its own data center revenue more than doubled. Third, the leader is running too. NVIDIA now brings out a new generation of chips about every year, and its newest, Vera Rubin, is expected to supply about a fifth of its data center revenue this quarter. The arch describes what is left of a lead after the chase has run its course. For NVIDIA the chase is still running, and the question is which kind of lead it holds.

Now set the price beside the chase. In the quarter that ended in July 2026, NVIDIA earned $63.7 billion from operations, about $213 billion a year after tax at that rate. On its book equity of $229 billion, at a cost of equity of 9%, that leaves about $192 billion a year of profit above what its capital costs: its excess profit. Held flat forever, never growing and never fading, that excess profit would make the company worth about $2.4 trillion, about 41% of its price. With the 21% average fade, it would be worth about $0.9 trillion; with the 38% base rate, about $0.6 trillion. For the price to make sense with no fade at all, the excess profit must rise to about $499 billion a year, 2.6 times today's, and stay there forever. Or it must grow about 52% a year for five years, to eight times today's, and only then begin to fade at the average rate.

That is not an implausible near future. NVIDIA's chief financial officer has told investors to expect revenue to grow about 70% in its fiscal year ending in January 2028. But the price is a bet on more than growth. It is a bet that NVIDIA's lead belongs to the second kind in the model, a lead whose copying cost rises with it: the CUDA software that most AI code is written for, the networking that ties its chips together, the racks it now sells whole, and $279 billion of commitments to the suppliers of the parts that are scarcest. Those are Teece's complementary assets. They are formidable. The custom chips say the largest customers can copy the part of the work that matters most to them, at a cost they can afford, for the jobs they run most often. The tests below will show which force is winning.

The lead nobody took

The lead that lasted longest in this inquiry belongs to a company most investors have never heard of. NVE Corporation, of Eden Prairie, Minnesota, makes spintronic sensors and couplers, many of them for medical devices. Its sales have ranged between $21 million and $38 million a year since 2010. In those seventeen years its operating margin has never been below 49.9% and has reached 67%; in the fiscal year that ended in March 2026 it was 60.5%. Its annual report says that in couplers its strategy is to compete on features rather than solely on price, and it claims for its medical sensors advantages of reliability, size and sensitivity; among its risks, it also warns that competition has meant "more competitors and more severe pricing pressure."

Its filing counts competitors "many of whom have significantly greater financial, technical, and marketing resources," and still none has taken its margin. By the base rate, its margin's excess over the average should have halved within a year and a half and all but vanished within a decade. By the model, it is exactly what a small prize looks like. The lead is real. The market is small. For the largest technology companies in the world, the reward for catching up is not worth the trouble of the chase. Persistence, here, measures the size of the prize as much as the quality of the lead.

What it means for an investor

Vista makes no buy or sell calls. The analysis still points in clear directions.

Judge a lead by the size of its prize and the cost of copying it, not by how long it has lasted. For any company with a valuable lead, ask how much that lead is worth to its best-funded rival, and whether copying it gets more expensive as it grows. A long record of high returns answers neither question.

Where the prize is enormous and copying costs about the same whatever the prize, as for AI models, drugs after their exclusivity and published investment strategies, expect the fade to run faster than any average, and value the lead by what it earns while it lasts. In the model's terms the value of a great lead is mostly its early earnings, the b term, and a price that depends on the late lead, the d term, is a price that bets against the chase.

The leads that last are of two kinds: small leads in markets too small to chase, and leads whose copying cost rises with them, because they rest on complex systems or on complementary assets a rival cannot buy. The first kind is easy to overlook. The second is easy to assume.

For NVIDIA the price assumes the second kind. For the companies that sell AI, the model says the gains go to whoever owns what the copier cannot copy. For the GLP-1 drugs, it says the new leader will be chased as the old one was.

The signposts, in order: AMD's results on November 3, 2026, and Novo Nordisk's on November 4; NVIDIA's results for the quarter ending in October, expected about November 18; Broadcom's for its fiscal fourth quarter, expected about December 10; Novo's annual report on February 3, 2027 and Lilly's results for the fourth quarter in early February; NVIDIA's and Broadcom's next reports, expected in late February and early March; and NVE's annual report, expected in early May 2027.

The data that we're missing is...

• How the largest cloud companies divide their work between NVIDIA's processors and their own chips, by kind of job, training or inference, and what each costs per unit of useful work. The filings give revenue, never the split.
• What it really costs to copy the frontier. DeepSeek's $294,000 excludes the research that came before it, and no lab publishes its full cost of following.
• Whether NVIDIA's price increases hold against the custom chips, and at which customers.
• How long it takes a large customer to move a workload off NVIDIA's software, and what it costs.
• How Lilly's and Novo's net prices in America are moving, under compounded copies, cash payment and the list price cuts of January 2027.
• Why NVE's niche stays unchased: how long a medical device maker takes to qualify a new sensor supplier, and how much each design is worth.

Primary research

Likely, we can get some of it from primary research, which is what Vista is all about. The people who planned the cloud companies' chips, built the models that copied the frontier, sold the weight-loss drugs and designed the sensors already know what the filings will take a year or more to show.

For a client engagement, Vista combines this report with that primary research.

The checkpoint, set down now

We will score this inquiry on May 31, 2027, and publish the result on the scorecard whether it flatters us or not. Each test sides with the reading here:

• The chase reaches NVIDIA. Broadcom's AI semiconductor revenue for its fiscal first quarter of 2027, ending about February 1, 2027, is at least 25% of NVIDIA's data center revenue for its fiscal fourth quarter of 2027, ending about January 25, 2027. It was 18.8% in the latest comparable quarters, and 11% to 14% in every quarter from 2024 through the first of 2026. Falsified if it is below 25%.
• The new leader is chased too. Lilly's share of American prescriptions for incretin drugs in the fourth quarter of 2026, as Lilly reports it from IQVIA data in its results presentation, is 62% or lower. It was 60.9% in the second quarter of 2026, after three quarters between 60% and 61%. Falsified if it is above 62%.
• The small lead holds. NVE's operating margin for its fiscal year ending March 31, 2027 is at least 50%, a level it has missed once in seventeen years, by a tenth of a point. Falsified if it is below 50%.

Three passes and the reading here gains ground; two is too early to say; one or none, and the chase does not work the way the model says.

On the front in Ukraine, the jammer that protected a trench in January 2024 was outmatched by that fall. Ford gave away the best technology of the century and was paid for it in fifteen million cars before the industry caught him. The Wrights kept theirs and were paid for it in lawsuits. Carroll's Red Queen runs as fast as she can to stay where she is, and so, it turns out, do the companies with the best ideas. The lead that lasts is not always the best one. Sometimes it is only the one nobody wanted.

Questions this inquiry answers

Can the best technology lose its lead faster than a lesser one?

Yes, when copying costs about the same whatever the prize. The more a lead is worth, the more a rival gains by catching up and the harder it tries, so the most valuable leads can keep the least of their advantage by a later date, even though they earn the most along the way. In a simple model of the chase, the catch-up arch, the lead that survives is largest for middling innovations. When copying gets more expensive as fast as the lead grows, the best keeps the most.

What is the catch-up arch?

A model of a rival deciding how hard to chase a leader. With a lead of size q and a prize-to-cost ratio k, the rival catches up with probability k times q, at most certainty, so the lead that survives is q times (1 minus k times q). It rises with the size of the lead, peaks at q = 1/(2k) and falls to zero at q = 1/k. With k = 1, a lead of 0.8 earns 50% more than a lead of 0.4 but keeps only two thirds as much of its advantage at the end.

How fast do competitors copy new products?

In Edwin Mansfield's 1981 study of 48 products, copying cost imitators about 65% of what the innovation had cost and took about 70% of the time, and about 60% of patented successes were imitated within four years; the details of new products usually leak to rivals within about a year. Richard Levin's 1987 survey found typical unpatented products copied within a year in two industries out of three, while major patented products took longer and could be copied by fewer firms.

Why did Novo Nordisk lose its lead in weight-loss drugs to Eli Lilly?

Novo's semaglutide had the lead first, with Wegovy approved in June 2021, and the size of the prize drew the hardest chase. Lilly's tirzepatide pulled even in sales in the second quarter of 2025 and passed it clearly in the third, about four years and four months after Wegovy's approval; by the second quarter of 2026 it sold 1.61 times as much and held about 61% of U.S. prescriptions for the class. Novo's market value fell from more than $600 billion in 2024 to about $165 billion.

Are rivals catching up with NVIDIA's lead in AI chips?

The chase is visible. Broadcom's AI chip revenue ran at 11% to 14% of NVIDIA's data center revenue through early 2026 and reached 18.8% in the second quarter; IDC reports that servers built on custom chips rose from about 11% to about 24% of spending on accelerated servers in a year; and Google, Amazon, Meta, Microsoft and OpenAI all have AI chips of their own. This inquiry's first test asks whether Broadcom's share reaches 25% in the quarters ending in early 2027.

What does NVIDIA's stock price assume about its lead?

At $239.24 on October 6, 2026 and a 9% cost of equity, NVIDIA's current excess profit of about $192 billion a year, held flat forever, would be worth about $2.4 trillion, 41% of its price. The price needs that excess profit to reach about $499 billion a year and never fade, or to grow about 52% a year for five years before fading at the average rate of 21% a year.