The Throne and the Asteroid

A throne is not lost when a challenger arrives. It is lost when the board it was built on disappears.

Flash did not lose to HTML5.

Flash was extraordinary. Interactive web, streaming video, browser games, rich animation — Flash did things that seemed impossible through any other means. Adobe was serious. The developer ecosystem was deep. The content library was vast. The stakeholder gravity was real enough that entire industries had organised around Flash's continued existence.

None of it mattered.

Steve Jobs wrote a letter. Mobile proliferation arrived on Apple's terms. The board Flash had optimised for — desktop browsers, keyboard and mouse, abundant processing power and no battery constraint — was not the board that the next billion users came in on. Flash's greatest strength on the old board, rich interactive media requiring significant compute, was fatal on the new one. The asteroid was not HTML5. The asteroid was the board shift itself.

Flash's stakeholders did not know the asteroid had landed until the content started disappearing. By then the juvenile recruitment had already rerouted. Developers building for iPhone were HTML5-native from the start. The Flash developer community was the last generation without knowing it.

Power-hungry things do not survive board shifts.

This is the Nvidia story. Not the story being told right now about sanctions evasion and gray channels and Jensen Huang at noodle shops with Chinese executives. That story is real and its legal consequences will be what they will be. But it is the surface story. The structural story runs deeper and longer and matters more.

Nvidia holds the most valuable throne in the world.

The throne is not the GPU. The GPU is the visible seat. The throne is CUDA — seventeen years of developer tooling, optimisation libraries, academic papers written assuming CUDA availability, institutional knowledge embedded in a generation of machine learning researchers, cloud infrastructure built around Nvidia's architecture, startups founded by people who never learned to build on anything else.

CUDA is the most successful installed early architecture in the history of computing infrastructure. Every PhD student who learned PyTorch on CUDA, every startup that built its first model on an Nvidia instance, every enterprise architect whose mental model of AI infrastructure is GPU-shaped — these are not customers. They are the board itself. They are what the throne is built on.

This is the distinction that almost all Nvidia analysis misses. The revenue is downstream of the installation. The market cap reflects the installation. The moat is the installation. Nvidia is not dominant because its GPUs are the best. Nvidia is dominant because the world's AI developers were installed on Nvidia during the critical window when installation was cheap and alternatives were absent.

That window is closing.

China said NO.

Not because US enforcement got better. Because Beijing made a strategic calculation that Nvidia dependency was a vulnerability that had to be unwound before it became impossible to unwind. The state-funded data center ban, the border blocking of H200 shipments, the directive to prioritise domestic alternatives — these are not reactions to American pressure. They are deliberate board redesign. China identified the dependency, assessed the exit cost, and chose to pay it while the domestic alternative stack was still immature enough that paying the cost was painful but survivable.

The gray channel that has consumed so much journalistic attention — the shell companies, the Malaysian subsidiaries, the dummy servers, the Southeast Asian pass-throughs — was not China cheating the system. It was China's transition infrastructure. The gray channel kept Chinese AI development running while the domestic stack matured enough to replace it. Now that the domestic stack has crossed the minimum viability threshold for most Chinese workloads, Beijing is shutting the gray channel from its own side. Not because the Americans forced it. Because China no longer needs it.

Huawei generated $7.5 billion in AI chip revenues in 2025. That number receives a fraction of the attention it deserves. A protected domestic market with $50 billion in annual AI compute demand is the most powerful development subsidy imaginable. Every dollar of Chinese AI spend redirected from Nvidia to domestic alternatives funds the maturation of those alternatives. The domestic stack is getting better because it has to. The board demands it.

China's exit from Nvidia dependency is not a revenue problem for Nvidia. It is a demonstration problem. China has proven to every government on earth that dependency on Nvidia is a strategic vulnerability and that exiting that dependency, while painful, is achievable. That demonstration is more dangerous to Nvidia's long-term position than the revenue loss.

The in-between markets are where the next decade gets decided.

Not the frontier. The frontier — the top AI research organisations, Western hyperscalers at the bleeding edge, the AGI labs — stays on CUDA. The switching cost is now genuinely prohibitive and the performance gap remains real. Nvidia owns the frontier completely and will continue to.

But the frontier is not the whole market. Most of the world's AI infrastructure by volume is not frontier training. It is inference. Recommendation systems. Computer vision. Language model deployment at enterprise scale. These workloads run adequately on hardware that is 70% of Nvidia's performance at 40% of Nvidia's cost.

China's domestic stack, developed under the forcing function of a massive protected market, will reach that threshold. It is probably already there for significant workload categories. When Chinese hardware vendors — Huawei and the entrants that China's protected market investment will produce — look for markets outside China to generate the scale that justifies continued R&D investment, the in-between markets are the obvious target.

Southeast Asia. India. Middle East. Africa. Latin America. These markets are growing AI infrastructure investment rapidly. They are price-sensitive. They are not doing frontier research. They are building exactly the infrastructure where Chinese alternatives are already competitive.

The financing relationships and political infrastructure that decades of Chinese engagement built in these markets are the distribution channel. The sale happens through the infrastructure relationship before any technology decision is made. A government that gets Chinese development financing for its data center build gets Chinese hardware as the default. The developers who build on that hardware get installed on the Chinese stack during their formative years.

This is the juvenile recruitment mechanism operating at geopolitical scale. The replacement ecosystem does not announce itself. It builds quietly in the markets the incumbent is not defending aggressively, installs itself into the next developer cohort, and reaches critical mass before the incumbent notices the gap.

Flash's replacement did not know it was Flash's replacement. It just built for the board that was there.

The hyperscalers are building the hedge.

Google's TPUs. Amazon's Trainium and Inferentia. Microsoft's Maia. Apple's Silicon trajectory. These are not products designed to beat Nvidia in the open market. They are designed to reduce dependency on a single supplier whose pricing power, geopolitical exposure, and supply chain concentration have become strategic liabilities for the companies most exposed to them.

When AWS optimises its own AI services on custom silicon and offers those services to customers, customers do not need to think about the underlying hardware. The Nvidia abstraction layer disappears. The developer builds on AWS primitives. The infrastructure choice has already been made above them. The juvenile recruitment happens through the platform without the developer ever making a conscious hardware decision.

This is the mechanism that matters most in the Western market. Not a competitor building a better GPU. A platform controller making a different default. The developer installs on whatever the default cloud instance runs. Change the default and you change the installation board for the next generation.

The hyperscalers are not saying NO to Nvidia today. They are building the infrastructure to say NO when the moment is right. The custom silicon exists. The developer tooling is maturing. The performance gap is closing for inference workloads. The moment when the default shifts is approaching.

When it arrives it will not look like a decision. It will look like an infrastructure update.

The NO event changes everything.

The asteroid metaphor is precise because the asteroid does not negotiate with stakeholder gravity. Flash had enormous stakeholder gravity. Every web developer, every content creator, every enterprise with Flash-dependent infrastructure — all of them had reasons to want Flash to survive. The board did not care. The asteroid landed and the ecological chain beneath it reorganised around the impact.

Nvidia's stakeholder gravity is larger than Flash's was. The developer ecosystem is deeper. The institutional investment is greater. The switching cost is genuinely higher. But stakeholder gravity does not stop asteroids. It only determines how long the debris field takes to settle.

The NO is not one event. It is a sequence. China's NO has already arrived. The hyperscaler custom silicon hedge is the NO being built. The EU's regulatory pressure is the NO being drafted. Each one individually is manageable. In sequence they represent the board shift.

When apex predators fight the ecological chain beneath them destabilises regardless of who wins. Every company, every developer, every government sitting beneath the US-China technology conflict is making hedging decisions right now. The hedge is the rational response to watching two apex predators fight over the board. Nobody wants to be fully dependent on either combatant when the outcome is uncertain.

The hedging is the juvenile recruitment of the replacement.

Nvidia's response to all of this is the platform transition.

The robotics bet. Isaac. Physical AI. The full stack play for autonomous systems and humanoid robots and industrial automation. Jensen's argument is that physical AI is the next compute frontier and that Nvidia is positioned to own it the way it owned the GPU-for-AI transition.

If this succeeds it is the most important corporate pivot since Apple's transition from computers to consumer electronics. A company that owns the platform for physical AI has a moat that is potentially deeper than the CUDA moat because the latency requirements and sensor integration demands of real-world robotics create genuine technical barriers that pure software alternatives cannot easily route around.

If this fails — if physical AI arrives slower than projected, or fragments across too many form factors for a single platform to dominate, or if the transition window closes before Nvidia completes the pivot — then the company that optimised so completely for data centre GPU dominance finds itself without a successor board to move to.

The race is platform transition speed versus NO arrival timing.

If the transition completes before the NOs land in sequence the throne shifts to new ground before the old ground disappears. Nvidia survives as a different company on a different board with a different moat.

If the NOs land in sequence before the transition completes the company that built the most dominant position in technology infrastructure in a generation finds itself defending shrinking ground with a platform transition half-finished.

Flash became a cautionary tale not because it was bad technology.

It became a cautionary tale because it optimised so completely for one board that it lost the capacity to adapt when the board changed. The power-hungry architecture that made it extraordinary on desktop made it impossible on mobile. The stakeholder gravity that made it dominant made it slow to acknowledge the shift. The installed ecosystem that made it irreplaceable made it complacent about the conditions under which it was irreplaceable.

Nvidia is not Flash. The analogy is not one-to-one. Nvidia has resources, engineering talent, strategic clarity, and a platform transition underway that Flash never attempted. The frontier market is real and defensible in a way Flash's position never was.

But the structural warning is identical.

Dominance built on a specific board configuration is permanent only as long as the board holds.

The board is shifting.

Not collapsing. Not disappearing overnight. Shifting. The way boards shift — slowly at the edges, then all at once at the centre.

The asteroid has not landed yet.

But the sky is different than it was.

And power-hungry things that cannot read the sky do not survive what comes next.

This is an old essay which I forgot to post before. Might not be updated with current affairs.