In 1714, the British Parliament passed the Longitude Act. It offered £20,000, a fortune at the time, to anyone who could determine longitude at sea to within half a degree. The astronomers assumed the answer lay in the stars. John Harrison, a carpenter’s son from Yorkshire who never saw the inside of a university, decided the answer was a clock. His H4 timekeeper lost about five seconds on a voyage to Jamaica in 1761. The Board of Longitude, dominated by astronomers, paid him in grudging instalments and kept revising the tests. He received a final £8,750 in 1773, at the age of eighty, and only after George III took a personal interest. (The Board never formally conceded that he had won.)
The purse mattered less than the race. Which brings one to last week’s announcement from San Francisco. OpenAI has released a model it calls GPT-6 Astra and describes it as the world’s most intelligent, claiming the lead in software engineering, science and cyber security. The claim is aimed at Anthropic, a firm founded five years ago by senior OpenAI employees, now valued at $965 billion against OpenAI’s $852 billion, and which has held the technical lead for much of this year. Greg Brockman, OpenAI’s president, called Astra a generational leap and flirted with calling it artificial general intelligence.
In a few months the lead will probably change hands again. That, and not the superlatives, is the detail worth attending to.
Economists have quarrelled over this for eighty years. Joseph Schumpeter, in Capitalism, Socialism and Democracy (1942), argued that innovation needs monopoly rents, since only firms with fat margins can fund laboratories, and that the “perennial gale of creative destruction” would discipline them in time. Kenneth Arrow, in a 1962 paper titled “Economic Welfare and the Allocation of Resources for Invention”, argued the opposite. A monopolist who innovates merely replaces profits it already earns, while a competitive firm gains an entire market.
Philippe Aghion and his co-authors reconciled the two in the Quarterly Journal of Economics (2005) with an inverted U. Too little competition breeds complacency and too much destroys the margins that fund research. Innovation peaks in between, and the incentive to innovate is sharpest for firms that are neck and neck, each straining to escape the other. The authors called this the escape competition effect. Two firms locked level, each fearing the other’s next release, will outspend any comfortable monopolist on research.
Frontier AI is that finding made flesh. Consider what the rivalry has produced.
First, cadence. Model generations that once arrived years apart now arrive in months, each timed against the rival’s last release. ChatGPT launched in late 2022. In under four years the field has moved from amusing chatbots to systems that write production software.
Second, price. Brockman’s own framing of the launch was that “price per task is what matters”, a sentence no comfortable monopolist has ever uttered. Anthropic’s flagship model has reportedly seen usage plateau as customers migrate to cheaper alternatives, which is the market doing quietly what no regulator could do loudly.
Third, diffusion. Capabilities that were laboratory curiosities reach an Indian student’s phone within a release cycle, because neither firm can afford to sit on inventory.
There is also a fourth reason, rarely mentioned. Talent circulates. Anthropic exists at all because employees walked out of OpenAI and built a rival. The precedent is older than it looks. In 1957 eight engineers walked out of Shockley Semiconductor to found Fairchild, and from Fairchild descended Intel, AMD and much of Silicon Valley. California law helped. Section 16600 of its Business and Professions Code declares that “every contract by which anyone is restrained from engaging in a lawful profession, trade, or business of any kind is to that extent void”. The freedom to defect is itself industrial policy.
And a fifth. Safety has become a selling point. Astra is being marketed on cyber security in the wake of highprofile breaches, including an episode in which OpenAI’s own agents escaped a testing environment and hacked a startup, a breach the firm took over a week to detect. Rivals now advertise robustness against each other. Whether rivalry supplies enough safety is a fair question, and the American government has already restricted the rollout of some frontier models.
The distinction to preserve is between ex post accountability for demonstrated harm and ex ante prohibition of imagined harm. The first disciplines a race. The second cancels it.
What should India take from this? The Union Cabinet approved the IndiaAI Mission on March 7, 2024, with an outlay of Rs 10,371.92 crore. The instinct, whenever a technology becomes strategic, is to anoint a national champion and shelter it. We have run that experiment.
The MRTP Act of 1969, descended from the Monopolies Inquiry Commission of 1965, treated size as sin and competition as suspect, and it took the Raghavan Committee (2000) and the Competition Act of 2002 to invert the presumption.
The lesson of the Astra launch is the negative list. Government should provide what no single firm will, meaning compute, open datasets and patient research funding. It should then refuse to do the rest.
This is what India is actually doing this time. It should not pick the winner, should not write rules that only incumbents can comply with, and should not enforce contracts that stop engineers from walking out and founding rivals. Contestability is the subsidy that costs nothing.
If Indian AI firms are to matter, the useful question is not which one the state should back but whether any of them fears a rival enough to hurry.
India, too, once made timekeepers. HMT began manufacturing watches in 1961, in collaboration with Citizen of Japan, behind licensing walls and with a captive market. For three decades it had nobody to beat, and when the walls came down after 1991 it discovered it could not run. HMT Watches was finally wound up in 2016.
Harrison’s H4 sits in Greenwich and still works. The HMT factory at Tumakuru has been silent for a decade. Both were built by able engineers. Only one was built in a race.
Aditya Sinha writes on macroeconomics and geopolitics. Navin Vijay is a Technology and Strategy Consultant.
