Inside the AI Chip Race: Why the World Still Runs on TSMC
- Sparknify
- 13 hours ago
- 7 min read
Updated: 6 hours ago
From Silicon Valley to Taiwan, investing in “hard tech” is no longer optional — it’s existential.
When Jensen Huang of NVIDIA touched down in Taiwan this month, he wasn’t just visiting a supplier. He was paying homage to the beating heart of modern computing. Standing before thousands of TSMC employees in Hsinchu, Huang declared:
“Without TSMC, there is no Nvidia today.”
It wasn’t hyperbole. In the age of AI, where every breakthrough model is measured in petaflops and trillions of parameters, the ultimate bottleneck isn’t algorithms — it’s wafers.
The Fabric of Modern Intelligence
TSMC (Taiwan Semiconductor Manufacturing Company) quietly powers the world’s most advanced chips — the ones that drive ChatGPT, Tesla’s autonomous systems, Apple’s iPhones, and the GPUs filling hyperscale data centers from Oregon to Singapore.
Founded in 1987 by Morris Chang, a visionary engineer who left Texas Instruments to launch an untested “pure-play foundry” model, TSMC made a radical bet: that designing and manufacturing chips could — and should — be separate businesses.
Nearly forty years later, that model has defined the global tech landscape. TSMC manufactures over 90% of the world’s leading-edge chips, including Nvidia’s Blackwell and Hopper GPUs, Apple’s A- and M-series processors, and AMD’s Ryzen and Instinct lines.
As Huang told employees during his November visit:
“Our business is very strong, and it’s growing month by month, stronger and stronger.”
The AI gold rush, it turns out, is a silicon rush — and Taiwan sits squarely atop it.
Elon Musk’s “TeraFab” Dream Meets Hard Reality
The latest headlines underscore the point. In late October, Elon Musk revealed that Tesla may have to build its own “TeraFab” — a dedicated chip fabrication facility — to keep up with the company’s insatiable AI demand.
“I think we may have to do a Tesla TeraFab. It’s like [TSMC’s] Giga[fab] but way bigger. I cannot see any other way to get to the volume of chips that we’re looking for.”
Jensen Huang, however, offered a sober counterpoint:
“Building advanced chip manufacturing is extremely hard. It’s not just building the plant — the engineering, the science and the artistry of doing what TSMC does for a living is extremely hard.”
So yes: it takes more than money to replicate TSMC’s position.
Why No One Can Replicate TSMC
It’s not just about money.
A modern fab costs between $20 billion and $40 billion, employs thousands of engineers, and operates at atomic precision — literally. Even nations struggle to replicate what Taiwan has achieved.
TSMC’s secret is not just scale — it’s integration. The company has spent decades perfecting a supply chain that spans:
Extreme Ultraviolet (EUV) lithography machines — each costing more than a 747 jet.
Sub-nanometer process nodes — 3 nm today, 2 nm and below tomorrow.
Advanced packaging and chiplet integration, allowing multiple dies to function as a single processor.
Yields that would bankrupt competitors — each wafer represents hundreds of thousands of dollars in value.
Even when nations offer incentives — like the U.S. CHIPS Act or Japan’s Rapidus initiative — what’s missing is the tacit knowledge built through decades of iteration. Every step, from chemical vapor deposition to plasma etching to photolithography, must work in perfect concert.
That’s why Samsung, Intel, and GlobalFoundries — the nearest competitors — still trail TSMC in leading-edge nodes. Samsung is pushing 2 nm production (and recently signed a $16.5 billion chip deal with Tesla), while Intel’s Foundry Services is struggling to catch up after years of process delays. But the market share numbers are stark: TSMC still holds roughly 70% of all advanced foundry output.
Where the Edge of the Envelope Lies
Inside a modern fab, progress is measured in nanometers — billionths of a meter — and in power density, thermal efficiency, and yield. The frontier today sits around 2 nm, and TSMC is already accelerating its roadmap.
TSMC has moved up its 2 nm production plans in Arizona due to “overwhelming demand from AI customers” and hinted at expanding its $165 billion U.S. investment.
The challenges at this frontier are staggering:
Quantum tunnelling and leakage currents that defy classical transistor behaviour.
Multi-pattern EUV lithography to draw features smaller than the wavelength of light.
3D stacking and chiplet architectures to overcome the slowdown of Moore’s Law.
Thermal management — advanced chips can dissipate more than 700 W of heat.
Supply chain fragility — even a single contaminated batch can halt production for weeks.
The phrase “the most complex object humans make” isn’t exaggeration. A single 300 mm wafer can require over 700 steps and take months to process.
TSMC’s Official Response: Stability Amid Tension
As global tensions rise, Taiwan’s foundry giant has found itself in the middle of geopolitics. The U.S. recently informed TSMC that the export waiver allowing its Nanjing fab to supply chips to Chinese customers would expire at the end of 2025.
In a statement to Mobile World Live, TSMC said:
“We remain fully committed to ensuring the uninterrupted operation of TSMC Nanjing and are in communication with the U.S. government to evaluate appropriate measures.”
Despite the uncertainty, the company’s revenue guidance remains robust, with continued growth driven by high-performance computing and AI demand. In essence: the chips must flow.
Analysis: Why Elon Musk Is Trying to Move Away from TSMC
Tesla’s shift away from relying exclusively on TSMC is not simply about one supplier falling short — it’s a strategic reframing of chip manufacturing by Tesla and Musk. Analyst Ming‑Chi Kuo outlines the motivations, and Musk’s own remarks frame the urgency.
Three Key Factors
Geopolitical concentration & supply-chain vulnerability
Kuo notes that Musk is “clearly aware of the concentration of advanced chip capacity in Taiwan.” TSMC’s advanced-node and packaging capacity in the U.S. is likely to remain under 10% of its total by 2030. Musk has voiced concerns that:
“Even when we extrapolate the best-case scenario for chip production from our suppliers, it’s still not enough.”
By owning the fab infrastructure, Tesla would reduce exposure to cross-strait risk, export controls, and third-party supplier prioritisation.
Vertical integration and performance customisation
As Kuo argues, Musk’s ambition goes beyond supply shortages — it is about full control of the AI chip supply chain. Tesla’s roadmap from AI5 to AI6 within a year (as outlined by Musk at the shareholder meeting) underscores this:
“I’m hopeful that we can within less than a year of AI5 starting production, we can actually transition in the same fab to AI6 and double all of the performance metrics.”
By building its own fab, Tesla could tailor process, packaging, interconnect, power–thermal trade-offs, and integration with its vehicles/robots more directly than relying on external foundries prioritising other major clients (e.g., Apple, Nvidia).
Scaling volume to match ambition
Tesla’s ambitions in autonomous driving, humanoid robotics (e.g., the Optimus program) and large-scale AI compute mean wafer volumes far above what standard foundry partnerships may reliably deliver. Musk described needing “a gigantic chip fab” because existing supplier capacity won’t meet Tesla’s projected demand.
Kuo emphasises Tesla’s position as a second-tier customer at TSMC compared with Apple and Nvidia, affecting priority and R&D support. Owning the fab infrastructure gives Tesla the ability to prioritise and control ramp timing, yield scheduling and product roadmap alignment.
The Hard Reality
While the motivation is logical, the execution is daunting. As Jensen Huang pointed out:
“Building advanced chip manufacturing is extremely hard. It’s not just build the plant, but the engineering, the science and the artistry of what TSMC does for a living is extremely hard.”
Tesla risks taking on decades of manufacturing know-how, multi‐billion‐dollar fabs, ramp risk, yield shortfalls, supply chain fragility and process integration — all in service of its AI hardware ambition. Even so, Musk appears convinced that “there is no other way” if Tesla is to scale at the rate it expects.
Implications for Investors and Manufactures
For investors: Tesla’s pivot signals that the manufacturing bottlenecks in the AI hardware ecosystem are real — and may create upstream opportunities (fab equipment, materials, packaging) and competitive risk to incumbent foundries.
For foundries like TSMC: this is a reminder that large customers may evaluate vertical integration when capacity or customisation bottlenecks appear — but they may still rely on foundry leadership to keep pace.
For hardware/hard-tech entrepreneurs: Tesla’s move underscores that supply-chain independence is becoming a strategic asset in AI/robotics/automotive — but reaching that independence is a rare and high‐risk undertaking.
The Strategic Imperative: Investing in Hard Tech
This is where investing in hard tech comes into focus.
Software can scale with talent and time. Hardware — particularly semiconductors, batteries, and robotics — scales only through physics, manufacturing, and capital. And therein lies both the challenge and opportunity.
Hard-tech investing means betting on the infrastructure of innovation: the physical systems that make digital revolutions possible. The semiconductor industry epitomises this — the ultimate fusion of materials science, mechanical precision, and national strategy.
Key truths for investors:
Barriers to entry are enormous — which means moats are durable.
Economic cycles are brutal, but secular demand (AI, EVs, 5G, data centres) keeps growing.
Talent scarcity defines winners; expertise compounds over decades.
Geopolitics is no longer a background factor — it’s central to valuation.
Hardware excellence drives exponential software value.
In short: investing in semiconductors isn’t just buying into a market; it’s underwriting the infrastructure of intelligence.
Why This Matters for Founders and Researchers
For startup founders and researchers working in hardware, robotics, clean energy, or AI compute, understanding the semiconductor backbone isn’t optional — it’s essential.
Every sensor, controller, accelerator, and robot you design ultimately runs on silicon someone had to fabricate. The next generation of deep-tech startups — from quantum chips to neuromorphic computing — will depend on the same manufacturing mastery that TSMC perfected.
And as the AI economy shifts from “training” to “deployment”, the bottleneck will shift from compute access to manufacturing capacity. That’s where investors, governments, and innovators must look next.
Join Us: “Investing in Hard Tech” — Live at CreaTV San Jose
To explore this frontier, Sparknify is hosting a live-studio recording of Silicon Valley Unplugged (SV Up):
🎬 Investing in Hard Tech
📅 November 18, 2025 | 6:00 PM – 8:00 PM
📍 CreaTV San Jose, 38 S Second St, Downtown San Jose
This episode will feature Nicolas Sauvage, President of TDK Ventures, in conversation with Sparknify’s Jillian Sun — exploring how leading investors evaluate and scale breakthroughs in AI hardware, semiconductors, energy systems, and robotics.
As we build bridges between Silicon Valley and Taiwan, this event also sets the stage for our next summit: Bridging Silicon Valley and Taiwan: Semiconductor & AI Synergies, on January 13, 2026, where global investors and Taiwan’s semiconductor leaders will converge.
If you’re building in hard tech — or investing in it — these are conversations you cannot afford to miss.
The World Runs on Atoms, Not Just Code
Elon Musk’s ambition to build a “TeraFab” and Jensen Huang’s gratitude to TSMC both point to the same truth:
Hardware is destiny.
The cloud may be global, but it rests on silicon born in cleanrooms. Every AI model, electric car, and data centre traces back to the art and science of lithography and wafer yield.
That’s why understanding — and investing in — the semiconductor ecosystem isn’t just a financial play. It’s participation in the defining industrial transformation of our time.
