Brains, Fabs, and Bottlenecks
A Counterpoint to “Skipping Nvidia Left Amazon, Apple and Tesla Behind”
Karl Freund’s latest piece, Skipping Nvidia Left Amazon, Apple and Tesla Behind, lays out the dominant narrative. Nvidia has become the standard, and the firms that dodge it pay the price. Apple stalled. AWS fell behind. Tesla gave up.
That story is true, but incomplete.
Because sometimes, even the winning bet carries risks many never see.
A Basket on a Fault Line
Putting all your faith in one chipmaker is like putting all your eggs in one basket. Only this basket sits on a fault line.
Karl tracks the money and the math. I’m more interested in the factories, the ports, the people who know how to make the chips that make AI possible.
When you pull those threads, the story gets interesting fast.
Nvidia’s dominance is real, but it’s brittle. Nvidia designs the microchips for its GPUs, but the chips themselves are fabricated in Taiwan by TSMC (Taiwan Semiconductor), only a hundred miles off the coast of mainland China.
TSMC recently built a new plant in Arizona, and they’re turning out wafers there now. But they’re already running into major problems with production.
This is where the story shifts. Nvidia’s dominance rests on an ecosystem that is more fragile than it looks.
Let’s turn this story over. We’ll start with the production hurdles, then the staffing challenges, and finally dig into the suppliers who supply the suppliers.
Karl is right on the quarterly scoreboard. It’s foolish to skip Nvidia. But my aim is to show just how delicate that bet really is.
Nvidia’s Golden Goose Still Lays in Taiwan
Nvidia’s dominance still depends on TSMC’s leading-edge fabrication facilities. The newest designs always debut in Taiwan first, and TSMC has already said its U.S. plants will not get the bleeding edge right away.
Arizona has started producing 4-nanometer wafers. These parts take sophisticated tools and unforgiving precision to make, and even a few Blackwell chips are rolling off the line there now.
Schedules are shifting under demand pressure. The second Arizona fab, once slated for 2028, is now aiming for early 2027 or even late 2026. A third fab, built for ultra-high precision N2 and A16 nodes, might begin as early as 2028, a full year ahead of plan.
Timelines stretch under the weight of demand. That’s real progress. But the sharpest edge of chipmaking still lives in Hsinchu, Zhunan, and Longtan.
Nvidia’s strength remains chained to TSMC’s timeline. Treat those timelines as weather, not scripture.
Chess Pieces
Meanwhile, Washington doesn’t see Taiwan as just an island. It sees a flashpoint.
The CHIPS Act, America’s subsidy program for domestic chipmaking, isn’t charity. It’s sovereignty insurance. The jobs are incidental. The real goal is leverage.
TSMC’s Arizona campus alone has pulled in up to $6.6 billion in grants and another $5 billion in loans for three fabs in Phoenix.
Onshoring chip production is a fire escape. Once the fabs and tooling sit on U.S. soil, Washington has levers it can pull. It can invoke the Defense Production Act to prioritize defense needs. It can tighten export controls to ration supply. These are tools it has already used in past shortages.
Beijing sees the same board. Taiwan is still the choke point for the world’s semiconductors, so China is sprinting to build its own supply chain. Subsidies for domestic fabs, recruitment drives for engineers, whatever it takes to stand up a parallel system.
Taiwan itself is the hot potato. Neither Washington nor Beijing can touch it without setting the table on fire, yet both treat it as the prize that defines the game.
These plays are more insurance policies than solutions. They buy time, not independence. Real fixes take decades.
The Silicon Shield
TSMC is no ordinary company. It’s the most important factory on Earth, and the world knows it.
Washington calls it a silicon shield: break the island, break global tech.
Investors treat it as ballast for portfolios.
Taipei frames it as national security.
The protections around their facilities make that clear.
TSMC fabs are hardened against quakes and storms, buffered with redundant water and power, and bolstered by overseas plants.
But protection cuts both ways.
The more the world leans on one company, the more precarious the system becomes. Trillions in market value and billions in fresh AI spending now rest on a single point of failure. Microsoft, Amazon, Meta, and Google alone poured nearly $70 billion into AI infrastructure last quarter.
A single rumor out of Hsinchu can move global markets.
A hiccup in production can ripple through model training schedules.
The shield may hold, but the world still shakes.
And no amount of protection erases the chokepoints inside.
The flow of chips still runs through fragile steps in the chain.
None of those steps is tighter than packaging.
From Wafer to Workload
Karl emphasizes speed to market, and I agree that matters. But wafers don’t ship themselves.
The real hurdle is packaging. It’s the step that takes bare wafers and turns them into finished GPUs by bonding processors and their high-speed memory into a single working unit. Without it, you don’t have an AI chip; you just have expensive silicon sitting idle.
It’s like baking the layers of a cake but then having to send them across the ocean to be stacked, frosted, and decorated. And if they aren’t aligned to the millimeter, the whole thing collapses. Until that last step is done, it isn’t ready for the table.
Nvidia has already warned that packaging, not wafer output, will be its next constraint. Blackwell parts depend on a newer, more complex process that demands more equipment, tighter cleanroom tolerances, and near-perfect die-to-die alignment. That capacity is still clustered in southern Taiwan, with most of it running through TSMC.
The result is a waiting line that stretches across the industry. Every major GPU vendor is competing for the same packaging slots. TSMC is racing to add benches in the Southern Taiwan Science Park, but the crowd is growing faster.
Arizona can turn out wafers, but it still can’t finish them. Those wafers get crated up and flown back to Taiwan for the final step. Until packaging spreads beyond the island, the world’s GPU supply lives and dies at a single bottleneck.
And even if packaging expands, another wall waits at memory.
Upstream Choke Points
High Bandwidth Memory (HBM) is the ultra-fast stacked memory that keeps AI chips fed.
HBM supply is controlled by SK Hynix, Micron, and Samsung. Orders are already booked solid through 2025, with a waiting line already stretching into 2026.
Nvidia leads the market, but it doesn’t control this lever. Every competitor, from AMD to Apple, depends on the same three suppliers. Different baskets, same bottleneck.
The same fragility shows up further upstream.
Every leading-edge chip depends on EUV lithography: light-beam machines that etch the tiniest circuits. They’re built by ASML in the Netherlands, with optics from ZEISS in Germany and light-source tech from the U.S.
One supply chain, stretched across three countries, feeds every advanced fab on Earth. If one link snaps, the whole AI industry stalls.
That’s the upstream risk. The downstream lesson is just as sharp: you can’t simply copy-paste a fab.
Copy Exact and the Limits of Cloning
Chipmaking runs on a principle called copy exact.
Every pump, every pipe, and every process has to match the original fab, or yields collapse. Miss a step, and you don’t get chips. You get scrap. One bad drum of photoresist can erase months of production. In 2019, a single contaminated batch wiped out 30,000 wafers. Half a billion dollars gone.
On paper, this duplication looks simple: build the same kitchen, get the same dish. In practice, moving from Hsinchu to Arizona showed how much of TSMC’s consistency depends on an ecosystem you can’t export.
TSMC ships ultra-pure sulfuric acid from Taiwan to Los Angeles, then trucks it to Arizona. It isn’t about price. It’s about purity. U.S. suppliers can’t meet the grade TSMC demands. Even Intel, a veteran chipmaker with fabs next door, queued up for the same shipments. Other chemicals make the same trip too, because the local supply chain simply isn’t there.
In Taiwan, every spare part and contractor sits within an hour’s drive.
In Arizona, there’s desert.
Fourteen suppliers have promised U.S. facilities, but it will take years to grow them. For now, Arizona runs on a supply chain that’s longer, thinner, and more expensive.
And those gaps don’t just slow things down. They drive prices up.
Those costs reflect fragility. Morris Chang, TSMC’s founder, said bluntly that chips in the United States could cost 50 percent more than in Taiwan. In Arizona’s case, closer to double. Higher labor costs, weaker supplier density, slower permits, it all stacks up. Subsidies cover some of it, but not all. Resilience is lovely, but it also shows up on the balance sheet.
But copy exact exposes another problem too. You can’t just replicate machines and materials. You need the people who know how to run them. And that’s where the real bottleneck begins.
The Talent Trap
World War II gave us the original “Paperclip” lesson, when the U.S. imported German rocket scientists but couldn’t import the industrial ecosystem behind them. Semiconductor know-how works the same way.
You can import engineers with visas, but you can’t copy the supplier relationships, the social ties, or the institutional memory they grew up inside. You can poach individual experts, but you can’t replicate decades of tacit knowledge.
That reality showed up quickly in Arizona. TSMC had to fly in hundreds of Taiwanese engineers and contractors to install tools and train local staff. Without them, the fab couldn’t be brought online.
Cultural friction and unfamiliar labor rules only added delays. What takes two and a half years in Taiwan stretched much longer in the desert.
Arizona has the same equipment, the same blueprints, even some of the same managers. But yields lag until experience catches up.
TSMC is training new engineers at Arizona State and rotating veterans across the Pacific, but you can’t shortcut expertise.
Yield needs experience. And experience takes time.
Beyond the Near-Term Win
Karl’s right about the near-term. Skipping Nvidia hurt Apple, AWS, and Tesla.
But betting everything on Nvidia builds a different risk. Its foundation is exposed: packaging still clustered in Taiwan, expertise that doesn’t travel, and supply chains for memory and lithography that choke everyone the same way.
“Speed to market” matters. But when everyone rides the same hardware, they stack risk on the same fault line. Nvidia has become both the technical standard and the financial darling. That amplifies exposure, not resilience.
History rhymes: IBM’s mainframes looked untouchable. So did Microsoft’s browsers. Nvidia’s advantage is real, but it isn’t permanent. One systemic shock, and the illusion of durability cracks.
Don’t mistake the train for the tracks. Nvidia is running full steam, but the tracks run through Taiwan. Monopolies feel safe until they stall.
The U.S. and China are spending billions to escape those exposed points of failure. But subsidies only work if they make the whole chain sturdier, not just one link. Right now, this is rollover capitalism: short-term convenience dressed up as permanence, quarterly speed traded for long-term sovereignty.
Resilience comes from diversity. Competition spreads risk, but it also sparks invention.
If we want AI to be more than a house of cards, we need to invest across compute, packaging, memory, lithography, and most importantly, people.
Dominance looks permanent right up until it breaks. Nvidia’s edge is no exception.
Enjoyed this piece?
I do all this writing for free. If you found it helpful, thought-provoking, or just want to toss a coin to your internet philosopher, consider clicking below and donating $1 to support my work.
Enjoyed reading this!
Thank you for bringing more light to this matter. I always feel uneasy about Nvidia because I know the deeper we dig, the darker it gets. The upstream supply chain deliberately gets left out of these conversations - and that’s by design.
I’m actually unpacking this in my upcoming capacity building workshop on AI and African-rooted monitoring and evaluation. One goal is helping participants understand that AI isn’t weightless. When you trace back the supply chain, you see that materials for Nvidia’s chips are often mined in ways that violate human rights. The DRC is a perfect case study for exploring these dynamics.