How KXCO Ontology Helps Understand Markets

01Why compute is the deciding factor

The modern era of AI has a clean origin point. In 2012 a neural network called AlexNet won the ImageNet competition, and it did so not because of a single mathematical insight but because its authors trained it on NVIDIA graphics processors at a scale that had not been practical before. Everything since has been, in large measure, a story of scaling compute. Training computation for notable models has grown by more than a trillion-fold in a few decades, from roughly 10¹⁸ floating-point operations for early models to more than 10²⁵ for frontier systems like GPT-4 and its successors.

The reason this matters is that model performance scales predictably with the compute budget behind it, a pattern that researchers at Stanford's HAI, OpenAI and DeepMind have documented repeatedly and that the field now calls the scaling laws. More compute allows larger models trained on more data, and those models are simply better across a wide range of tasks. In practical terms, whoever controls the most computing power holds an outsized advantage in the race to build the most capable systems. That is why compute has also become a national-security question, and why the United States has restricted exports of advanced chips to China since October 2022, tightening the rules repeatedly through 2023, 2024 and 2025. The January 2025 AI Diffusion Rule went further, reaching closed-weight models trained with 10²⁶ operations or more. The unstated objective is plain: keep America's compute advantage, and with it its AI advantage.

The Stanford HAI 2026 AI Index framed the stakes precisely. The United States hosts 5,427 data centers and consumes more energy for computing than any other country. China has narrowed the model-performance gap to near-parity, an impressive result given its constrained access to the best hardware. But the foundational layer, raw compute capacity, remains decisively American, and that gap is widening as US technology companies pour hundreds of billions of dollars into new GPU data centers at a pace with no historical precedent.

02The global compute balance

As of mid-2025 the United States commanded roughly 75 percent of global GPU cluster performance, with China a distant second at around 15 percent and the remaining 10 percent spread across Europe, Japan, Singapore and elsewhere. Those figures come from a comprehensive dataset of AI supercomputers published in June 2025. A separate GeoCoded report in August 2025 put the US share of global public compute capacity at 48.4 percent, a lower number that reflects a different methodology but tells the same story of American dominance.

In absolute terms the United States is estimated to hold between 100 and 200 exaFLOPS of aggregate data center compute, and it is growing fast. AI-related power capacity is projected to jump from roughly 30 gigawatts in 2025 to 90 gigawatts or more by 2030, compounding at about 22 percent a year, with over 700 data centers under construction across 38 states, representing some 18 gigawatts of new capacity. China reports a much larger headline number, roughly 2,185 exaFLOPS of national intelligent computing capacity by mid-2025, but that figure uses a broader definition of compute than dedicated AI training clusters; its national data centers provide around 230 exaFLOPS.

Export controls are the reason the gap holds. Since October 2022 the US has banned exports to China of any AI chip at or above the capability of NVIDIA's A100, expanding the net in later years. The Trump administration briefly halted AI chip exports to China entirely in April 2025 before reversing course in July 2025 and approving sales of NVIDIA's H200. The framework remains restrictive. Chinese semiconductor firms delivered 1.65 million AI GPUs in 2025 out of roughly 4 million units in the country, taking 41 percent of local AI server shipments; the rest came from NVIDIA and AMD through restricted channels or older products below the control thresholds. The practical effect is that Chinese developers work with less capable hardware, rely on smuggled advanced parts, or invest in domestic alternatives a generation or two behind. Software cleverness narrows the gap. It does not close the physics.

03The United States: the compute ecosystem

Chip designers

The centre of gravity is NVIDIA (NASDAQ: NVDA). Founded in 1993 and based in Santa Clara, it designs the GPUs that power most AI training and inference worldwide, and its H100, H200 and forthcoming B200 and Blackwell Ultra parts set the global standard. Its market capitalisation of roughly 4.84 trillion dollars as of July 2026 makes it the most valuable semiconductor company on earth and one of the most valuable companies of any kind. Major institutional holders are Vanguard, BlackRock, State Street and Fidelity. NVIDIA has itself become a significant investor in other chipmakers, notably taking a large stake that makes it one of Intel's biggest shareholders.

AMD (NASDAQ: AMD), at around 250 billion dollars, is NVIDIA's main American rival, with its MI300X and forthcoming MI400 Instinct accelerators winning share among cloud and enterprise buyers who want a second source. Intel (NASDAQ: INTC), around 435 billion dollars after a strong quarter, is rebuilding around its own accelerators and foundry. Beyond the three giants, Broadcom (AVGO), near 1.82 trillion dollars, designs custom AI accelerators for hyperscalers including Google and Meta; Marvell (MRVL), around 265 billion; and Qualcomm (QCOM), around 180 billion, is pushing into data center AI with deployments confirmed by Meta and Microsoft. Cerebras (CBRS), which IPO'd in May 2026 at roughly a 26.6 billion dollar valuation, builds wafer-scale systems.

Hyperscale cloud providers

The true scale shows up in the clouds. Alphabet (GOOGL), above 2.4 trillion dollars, builds its own Tensor Processing Units alongside vast NVIDIA fleets. Amazon (AMZN), around 2.4 trillion, runs AWS, whose 2026 revenue is projected near 163 billion, and develops Trainium and Inferentia silicon. Microsoft (MSFT), above 3.3 trillion, runs Azure and is OpenAI's exclusive cloud partner, committing roughly 80 billion dollars to AI infrastructure in fiscal 2026. Meta (META), around 1.7 trillion, plans to deploy seven gigawatts of compute in 2026 alone. Apple (AAPL), around 3.5 trillion, is building AI server capacity for its own features. A newer category of specialist GPU clouds has also emerged, led by CoreWeave (CRWV), whose largest shareholder is NVIDIA, alongside Lambda, Vast.ai and infrastructure players like Applied Digital.

04China: the rising challenger

China's ecosystem is anchored by a company that is not listed anywhere. Huawei, founded in 1987 and employee-owned, generates over 100 billion dollars in annual revenue, and its HiSilicon subsidiary designs the Ascend series, China's closest domestic answer to NVIDIA's data center GPUs. The Ascend 910B and the forthcoming 910C are the leading edge of indigenous Chinese AI silicon. Huawei also runs a cloud, operates data centers and builds the servers and networking around them. Its constraint is fabrication: cut off from TSMC's most advanced nodes by sanctions, its chips are made at China's SMIC on older processes, which means lower performance and higher cost than NVIDIA parts built on TSMC 4nm and 3nm.

Among listed names, Cambricon (Shanghai STAR Market: 688256) is the standout, often called China's NVIDIA. Its market capitalisation sits near 780 billion yuan, roughly 103 billion dollars, after a stock that more than doubled in 2025 and rose again into 2026 on AI chip demand and its role supplying firms like DeepSeek. In cloud, Alibaba (BABA / 9988.HK) leads the domestic market with around 36 percent share and designs its own chips through T-Head; Tencent (0700.HK) is second; and Baidu (BIDU / 9888.HK) runs the fastest-growing cloud and the ERNIE model family. Other names, Sugon, Biren, Enflame, Moore Threads and ByteDance-backed Kunlunxin, round out a field that is broad but held back by the same manufacturing ceiling at SMIC.

05The foundry chokepoint

No account of this race is complete without the company that sits underneath all of it. Taiwan Semiconductor Manufacturing Company (NYSE: TSM) is the world's largest and most advanced contract chipmaker, producing roughly 90 percent of the world's most advanced semiconductors. Almost every leading AI chip, whether from NVIDIA, AMD, Broadcom, Qualcomm or Apple, is fabricated in TSMC's Taiwan foundries on its 3nm and 4nm nodes. Its market capitalisation of about 2.03 trillion dollars as of July 2026 makes it one of the most valuable companies in the world, and its position makes it the highest-leverage single node in the entire AI supply chain.

It is also its single largest point of failure. The concentration of leading-edge manufacturing in Taiwan is a vulnerability that both Washington and Beijing understand acutely. The US has pushed TSMC to build fabs in Arizona, now operational but at limited capacity next to the Taiwan plants. China's inability to reach TSMC for advanced work is among the most consequential effects of the cross-strait situation and export controls. SMIC, China's best domestic foundry, can reach the 7nm node and below using multi-patterning on older lithography, but its yields and performance trail TSMC's leading edge, which keeps a persistent bottleneck under every Chinese chip designer.

Value and fragility are often the same fact seen from two sides. TSMC is the clearest example on the entire board.

06The same board, drawn as an ontology

Everything above is accurate, and everything above is also hard to hold in your head at once. That is the problem an ontology solves. Instead of describing the compute race in paragraphs, the KXCO ontology represents it as a graph: every company, chip, foundry, person and government is a typed node, and every relationship, who fabricates whom, who owns a stake in whom, who is cut off from whom, is an edge with a source attached. Once the landscape is a graph, the things that matter to an investor stop being buried in prose and become visible as shape.

The map below is a compact view of the compute board drawn in that style. Press Run scan and it fans out from the ontology at the top, the way the live map reasons across the sector, and reaches the American cluster, the Chinese cluster, the foundry chokepoint in the middle, and the ranked findings at the bottom that flag where opportunity and risk actually sit. Hover any node to inspect it. This is the same structure that kxco.ai/ontology-live maintains across more than 220 nodes and 500 sourced edges; the version here is trimmed to the compute story so it stays readable.

Interactive. Drag nodes, hover to inspect, press Run scan to trace the ontology across the board. Rings appear on each node as the scan reaches it, marking risk findings, opportunity findings and the foundry chokepoint. Rendering is illustrative and trimmed for readability; the full live graph with sources is at kxco.ai/ontology-live.

07Reading opportunity and risk from the structure

Here is the investor use case in plain terms. The KXCO ontology does not just store the companies; it surfaces ranked findings, each marked as opportunity or risk and each backed by evidence nodes you can click through to sources. When you point that machinery at the compute race, a handful of findings do most of the work. They are worth stating directly, because each is an investment thesis and its own hedge at the same time.

Risk: a single point of failure at the top

The clearest risk finding is concentration. One company, NVIDIA, sits upstream of nearly every serious AI effort on the American side, and a four-trillion-dollar valuation encodes an assumption that its position holds. The ontology makes the dependency visible as a fan of inbound edges: hyperscalers, specialist clouds and enterprises all point back to the same supplier. Concentration like that is a source of extraordinary pricing power, which is the bull case, and a source of extraordinary correlated risk, which is the reason to size the position with your eyes open.

Opportunity: value sits upstream, at the chokepoints

The mirror-image opportunity finding is that in a supply chain this concentrated, the durable value tends to accrue upstream, at the narrowest points. TSMC is the canonical case. It captures a share of the economics of every advanced AI chip on the board regardless of which designer wins the season, because they all have to pass through its fabs. Broadcom and Marvell occupy a similar structural position in custom silicon and interconnect. The ontology surfaces this as the observation that the chokepoints, not the brand names, are where the leverage compounds.

Risk: the state now gates the frontier

A third finding is that government has moved from spectator to gatekeeper. Export controls, the AI Diffusion Rule, entity-list designations and the on-again off-again H200 approvals mean that policy, not only product, now moves these stocks. An investor who models NVIDIA or TSMC without modelling Washington and Beijing is missing a variable that has already whipsawed the group more than once. The ontology captures this as a set of government nodes wired directly into the companies they constrain.

Risk: a second frontier outside Western oversight

The fourth finding is that China is building a parallel, largely sanction-proof stack around Huawei Ascend and Cambricon, running its own models such as DeepSeek on domestic silicon. For a Western investor this is both a risk to the incumbents' addressable market and, through the listed names on the STAR Market and in Hong Kong, an opportunity set of its own, one that carries a different and heavier basket of political and disclosure risk. The point of the ontology is not to tell you which way to lean. It is to make sure the second frontier is on your map at all.

08The company tables

For reference, the two tables below collect the principal listed and private names on each side, with approximate market capitalisations as of July 2026. Figures are approximate and move constantly; they are drawn from company filings and market data and are provided for orientation, not as targets.

United States and Taiwan

China

09What we claim, and what we don't

10Frequently asked questions

Who is winning the AI compute race?

On the foundational layer of compute, the United States, by a wide and in many respects widening margin: roughly 75 percent of global GPU cluster performance to China's 15 percent, 5,427 data centers, and hyperscaler capex in the hundreds of billions. China has reached near-parity on model performance but remains a hardware generation or two behind because of export controls and its lack of leading-edge fabrication.

Which single stock matters most?

Two, for different reasons. NVIDIA is the most valuable and most depended-upon designer. TSMC is the most load-bearing, because it fabricates almost every advanced AI chip regardless of which designer wins. The ontology treats the second as the higher-leverage structural node even though the first has the larger market cap.

How does the KXCO ontology help an investor specifically?

It converts the sector from a story into a graph of sourced, typed facts, then surfaces ranked findings marked opportunity or risk. That lets you see concentration, dependency and hidden linkage directly, which is where opportunity and risk actually live, rather than inferring them from prose.

Can I see and check the live map?

Yes. It is public at kxco.ai/ontology-live, every claim carries a source link, and it includes an analyst-outlook view and an intelligence-findings view built for exactly this kind of reading.