For the last decade, the corporate world has been hypnotized by a single metaphor: “Data is the new oil.”
We believed it. We acted on it. Companies spent billions constructing massive data lakes and recruiting armies of PhDs to dredge them. The strategy was simple: whoever hoards the most proprietary information wins the future.
We were looking at the wrong map.
This matters to you if you’re building products powered by AI, running inference at scale, or competing in any market where machine learning creates competitive advantage. If your technology stack still treats “the cloud” as an infinite resource you rent by the hour, keep reading. You’re exposed in ways your CFO hasn’t modeled.
While we obsessed over our customers’ digital exhaust, the tectonic plates of technology shifted. As Paul Scharre details in Four Battlegrounds, the strategic advantage has moved from the software layer to the physical layer. The era of “Big Data” is ending. The era of “Scarce Compute” has begun.
The uncomfortable truth: Data saturates the market. Anyone can copy it. My use of a dataset doesn’t prevent you from using the same one. But compute—the physical hardware required to process that data—remains finite, rivalrous, and rapidly becoming the single most dangerous bottleneck in the global economy.
The Math Demands More Than Physics Can Deliver
From 2010 to 2022, the amount of compute used in cutting-edge machine learning research projects increased ten billionfold. Compute usage doubles every six months. Moore’s Law, which guided chip development for decades, doubled performance every two years. We’ve blown past that pace by a factor of four.
This creates a predictable squeeze: AI’s hunger for processing power outpaces the laws of physics that historically drove chip improvements. When OpenAI trained an algorithm to achieve superhuman performance at Dota 2, they used “thousands of GPUs over multiple months”—the equivalent of a human playing for 45,000 years. Training a robotic hand to manipulate a Rubik’s cube took 13,000 years of simulated computer time.
The money follows the math. Leading AI labs at OpenAI, DeepMind, and Google Brain spend millions on compute chasing the latest advances. DeepMind lost nearly $650 million in 2019 and had a $1.5 billion debt waived by Alphabet. Microsoft invested $1 billion in OpenAI in 2019. These numbers only work because the labs have backing from some of the world’s largest corporations.
You’re competing for GPU cycles against companies with deeper pockets and longer contracts. When capacity tightens, they prioritize their own products. You get what’s left.
The Supply Chain Runs Through a Minefield
The semiconductor supply chain is the most fragile industrial system in human history. It relies on a level of precision that makes aerospace engineering look simple by comparison.
Consider what it takes to produce a modern AI chip. The design typically comes from the United States—firms like NVIDIA, Qualcomm, and Broadcom dominate chip design, accounting for 65 percent of the global fabless market. But designing chips and manufacturing them are different games entirely.
The actual fabrication happens overwhelmingly in Taiwan. Taiwan Semiconductor Manufacturing Company (TSMC) alone accounts for over half of the global pure-play foundry market. Combined with other Taiwanese foundries, Taiwan controls 65 percent of chip fabrication globally. Raw market share understates their dominance. TSMC manufactures over 90 percent of leading-edge chips at the 7 nanometer process node and below. For AI chips specifically, TSMC produced eight of the ten AI-specialized chips available at advanced process nodes as of 2020, including the leading GPUs from AMD and NVIDIA.
Scharre puts it plainly: Taiwan is the Saudi Arabia of compute.
The precision required to make these chips depends on extreme ultraviolet (EUV) lithography tools—massive lasers that carve patterns smaller than a virus onto silicon wafers. Exactly one company in the world makes these machines: ASML, a Dutch firm. Without EUV lithography, you cannot manufacture leading-edge chips. Period.
Then there’s photoresist, a specialized chemical used in chip fabrication. Japan controls roughly 90 percent of the global supply for high-end EUV photoresist. When Japan had a trade dispute with South Korea in 2019, it temporarily restricted exports. South Korean chip manufacturers scrambled. Switching suppliers takes months of testing and recertification. There are no fast substitutes.
This isn’t a diversified global marketplace. This is a single-file line through geopolitical chokepoints.
Governments Weaponized the Chokepoints First
The United States understood something most businesses still haven’t internalized: while you can’t stop data from crossing borders, you can stop the machines required to process it.
Mark Leonard describes it in The Age of Unpeace: “As trade and value chains become so much more globalized, they have discovered that hitting one small link in the chain—such as chips or semiconductors—can be enough to bring a company or country to its knees.”
The U.S. government used export controls to strangle Huawei, China’s tech champion and global leader in 5G wireless technology. When the Commerce Department added Huawei to the Entity List in 2019, the ban initially had weak leverage. Within weeks, U.S. chip firms found legal workarounds and kept shipping chips. The problem was globalization itself—companies could sidestep bans by adjusting supply lines to cut out U.S.-origin technology.
So the U.S. government tightened the screws. It pressured the Netherlands to block ASML from selling EUV lithography machines to China. It expanded export controls to include any chip manufacturing equipment at the 16nm node and below. In September 2022, National Security Advisor Jake Sullivan announced a shift in U.S. policy: instead of keeping China “a couple of generations behind,” the new goal is maintaining “as large of a lead as possible.”
These aren’t tariffs. This is strategic amputation.
China imports roughly $300 billion worth of chips annually. It accounts for 60 percent of global semiconductor demand but struggles to produce advanced chips domestically. Every AI expert Scharre spoke with in China was acutely aware of China’s hardware gap. The U.S. export controls didn’t just slow China down—they severed access to the physical infrastructure required to compete in AI.
If your AI strategy relies on “cloud capacity,” you’re relying on a supply chain anchored in the most contested geography on Earth. You’re exposed to price shocks, rationing, and availability blackouts that will make the 2021 car chip shortage look minor.
What Hardware-Aware Strategy Actually Requires
Tech giants understand this. They’re engaging in a desperate, behind-the-scenes scramble to secure custom silicon. They’re buying their way out of the bottleneck. Most other businesses are merely renting space on a shrinking timeline.
When the next supply shock hits, who gets priority access to the GPU clusters? The cloud provider who owns them, or you?
Your technology strategy needs to account for scarcity the same way your supply chain team accounts for raw material availability. Here’s what that requires:
Map your critical path to specific hardware classes
- Stop forecasting cloud costs in dollars. Forecast them in GPU hours for specific chip architectures. If your product depends on NVIDIA H100s or A100s, and those chips become subject to allocation limits or export restrictions, which features fail? What happens if your cloud provider faces a 30% capacity reduction due to geopolitical tension or supply chain disruption? Walk through the scenario. What breaks? What becomes expensive enough to kill your unit economics? Run this exercise quarterly, not annually.
- Lock in capacity contracts while you can
Reserved instances and committed use contracts look expensive until spot markets double overnight. For mission-critical AI workloads, guaranteed capacity functions as insurance against supply shocks. Price the cost of not having access to compute when you need it. The 2021 semiconductor shortage should have taught us something: just-in-time supply chains work until they catastrophically don’t. - Evaluate partnerships and acquisitions through a hardware lens
A startup with brilliant algorithms but no secured compute access has fragile value. A team with proven GPU cluster allocations—even if their models are merely good, not great—might be worth partnering with or acquiring for the hardware guarantee alone. Look at every major vendor relationship and ask: what’s their compute situation? Are they subletting capacity from a hyperscaler who might prioritize their own products? Do they own their inference infrastructure? Can they maintain service levels if capacity tightens? - Build architecture flexibility to avoid single-vendor lock-in
If your entire AI stack runs exclusively on CUDA (NVIDIA’s software platform), you’ve locked yourself to a single hardware vendor. When NVIDIA raises prices or faces supply constraints, you have zero negotiating leverage. Investigate alternative frameworks like ROCm or OpenCL. Build some components on hardware-agnostic platforms. The engineering time costs you now, but it protects you from monopoly pricing and supply allocation decisions you can’t control. - Elevate compute supply chain risk to board level
This isn’t IT procurement. This is business continuity. Your ability to compete in 2026 depends on your ability to secure physical infrastructure that everyone else wants. A McKinsey study found that 97 percent of deep learning training in data centers used GPUs as of 2017. That percentage has only increased. Model the scenarios where compute becomes constrained, rationed, or prohibitively expensive. Assign executive ownership. Build contingency plans. Track geopolitical developments in Taiwan, semiconductor export policy, and fab capacity expansion.
The Inversion
Leonard describes the shift in geopolitics: “In the 20th century the world’s biggest economic choke-point involved oil being shipped through the Strait of Hormuz. Soon it will be silicon etched in a few technology parks in South Korea and Taiwan.”
Data is sand. It covers the beach. You can scoop it up by the bucket and still leave plenty for everyone else.
Compute is the furnace. There are only so many furnaces. They take years to build and billions to operate. The companies and countries that control them decide who gets to make glass and who gets to keep staring at sand.
Stop counting grains. Start securing your furnace.
