The end of the GPU arms race

Compute is becoming a utility. Context is not.

That single asymmetry is the whole argument, and it is worth stating before the evidence: anyone with capital can rent accelerators or call a frontier model's API. Almost nobody can hand that model an accurate, sourced, time-aware map of the domain it is being asked about. The first is a purchase. The second is an asset.

01 · SummaryKey facts

02 · The problemThe compute wall is not a hardware problem

The spending is not slowing. Amazon, Alphabet, Meta and Microsoft together guide to roughly $725bn of AI capital expenditure in 2026, up about 77 per cent on the prior year. Alphabet alone spent $44.9bn in a single quarter, double the year before. Microsoft has guided its next fiscal year to $255bn to $260bn.

What is changing is the market's willingness to pay for it in advance. In the final week of July 2026 more than a trillion dollars came off the chip complex. Nvidia lost $238bn of market value, SK Hynix $176bn, Samsung $173bn and Micron $113bn, with AMD and TSMC each shedding more than $100bn. The stated cause was not weakening demand. It was a repricing of expectations, on the worry that infrastructure spending is peaking faster than the revenue behind it.

Underneath that repricing sits a quieter technical fact. Scaling a language model improves fluency, breadth and reasoning on problems that resemble its training distribution. It does not create information the model never had. And a great many of the questions institutions actually need answered are exactly of that kind:

• Which of my positions route, at three hops, through a single lithography vendor?
• Which of my counterparties share a funding source I have not noticed?
• When did this person's stated position change, and what did they say before?
• Which of my suppliers became subject to an export rule, and on what date?

None of those is a language problem. Each is a traversal over a structure that either exists or does not. Ask a frontier model and you will get a fluent, confident, plausible answer assembled from whatever was in the training data, with no way to tell which parts are load-bearing. Ask a graph where every edge carries a source and a date, and you get an answer you can walk backwards.

This is not an argument against large models. It is an argument about where the marginal dollar goes. Past a certain point, the return on another order of magnitude of compute is smaller than the return on writing the domain down properly.

03 · The unitArchitected truth vs statistical guessing

An ontology, in Tom Gruber's formulation, is an explicit specification of a conceptualisation. Stripped of the academic phrasing, it answers three questions: what things exist, how they are related, and what that means. Andreas Kollegger of Neo4j put the practical case bluntly at the AI Engineer World's Fair in 2026: much of AI engineering is sliding toward controlled language, because one wrong word can send an agent sideways.

The smallest unit in KXCO's ontology is therefore not a company. It is a claim. Every relationship carries:

Two of those fields do most of the work. disc and conf mean a weak claim can be carried honestly rather than dressed up or dropped. A claim marked reported at medium confidence is not the same object as one marked filed at high, and an analyst who cannot see the difference is working blind.

The concrete effect: the single most important finding in the public map is not about a model company at all. It is that the sector resolves to one Dutch firm in Veldhoven making the extreme ultraviolet lithography every leading-edge AI chip requires. That firm was not in the seed set.

The instrument does not make the discovery. It makes the discovery possible.

It matters how that finding actually happened, because it is the thesis in miniature and it is easy to describe wrongly. The engine did not conclude that ASML mattered. It grew the graph outward from the companies everybody already discusses, laid the dependencies out visually, and the convergence became obvious enough that a person looking at the display recognised what it meant.

Every fact involved had been sitting in public filings for years. No datum was secret. What was missing was a rendering in which the shape of the thing could be seen at all. Without the display, nobody goes looking, and the most important structural fact about a trillion-dollar sector stays technically public and practically invisible.

This is the distinction that separates an ontology from the oracle people expect when you say AI. An ontology is an instrument, not an answer. It does not hand down conclusions; it puts structure in front of human judgement in a form that judgement can actually operate on. The discovery is always the person's. The contribution of the system is that it made the person's insight reachable.

Which is also why more compute does not substitute. A larger model asked about the AI supply chain returns a fluent paragraph, and a fluent paragraph is not a shape you can see the convergence in. No language model produces that finding as a load-bearing conclusion, because it is a property of the graph rather than of the text, and because being load-bearing requires a human deciding it is.

04 · TimeThe temporal trump card

Language models are stateless about time in a specific and damaging way. Asked about a person's position on something contested, a model averages across everything it absorbed, producing a blend of views held at different moments as though they were one settled opinion. In markets, where the sequence is the signal, that is worse than no answer.

KXCO's schema is bi-temporal. Valid time records when a thing was true in the world. Record time records when we asserted it. When something changes we do not edit the old claim; we close it off and add a new one that supersedes it, and the old one stays in the graph marked as history.

A worked example from the live map. Unsupervised vehicle autonomy has been promised on three separate recorded dates:

The third statement came less than a month after the second, with no reported resolution of the obstacles the second had named. What exists on the road today is fewer than forty unsupervised robotaxis across three Texas cities.

Notice what the structure gives you that prose cannot. The stance does not decay monotonically; it oscillates. A system that stored "opinion on autonomy" as a property of a person would have thrown away the interesting half. A system that stores each statement as a dated event, with the later superseding the earlier, lets you query the sequence, measure the gap between stated and delivered, and do it without anyone having to remember.

The same mechanism applies to policy, where it matters even more, because rules have exact effective dates. An export rule is not true. It is true from a date. When the US Bureau of Industry and Security moved the Nvidia H200 and AMD MI325X from presumption of denial to case-by-case review on 15 January 2026, that was an event with a timestamp, and every dependency downstream of it changed on that day and not before.

05 · FrictionContradiction and opacity as first-class data

Frontier models are trained to be agreeable, which makes them smooth over exactly the friction that carries the most information. Two disciplines matter here, and KXCO handles both deliberately.

Contradiction is computed, never asserted

It is tempting to store a CONTRADICTS edge between two people whose views clash. KXCO does not, and the reason is the same reason the rest of the map is trustworthy: every edge in the graph has a source, and our own inference is not a source.

Instead, stance is recorded on each dated statement, and opposition is derived at query time from statements that are still valid. The rule is visible, so it can be argued with. And it produces results that a stored edge never would.

Unknowns stay visible

No amount of compute makes a model admit the shape of what it does not know. A graph can, and the discipline is simply that a gap is left empty rather than filled. Of the 580 claims in the public map, 515 carry a source URL. The remaining 65 do not, and they are visible as such rather than dressed in a citation that does not support them.

Being straight about the current state: the schema supports a first-class opaque-dependency relationship, for cases where a dependency almost certainly exists but nobody has published it, and the public map does not yet use it. The refusal to fabricate is real and enforced. The machinery for marking a specific suspected-but-unpublished link is defined and not yet populated. Those are different things and it is worth not conflating them.

This is the hardest discipline to hold and the one that most distinguishes an intelligence asset from a content product. It is always tempting to fill a gap with a plausible estimate. A visible unknown is worth more, because a visible unknown can be closed. A fabricated estimate quietly poisons every conclusion downstream of it, and you will not know which ones.

06 · LayersSilicon, sanctions and theatres

Structure earns its keep when a layer you added for one reason answers a question you did not ask.

The public map carries silicon as first-class entities rather than as text in a note: the Nvidia H200 and H20, AMD's MI325X, Huawei's Ascend 950PR and 910C, Nvidia's Vera Rubin platform. It carries policy instruments with their effective dates: the BIS rule of January 2026, the US Entity List, China's AI-generated content labelling measures in force since September 2025, the anthropomorphic-services measures effective 15 July 2026, the EU AI Act. And it carries theatres, the places this technology is actually being tested.

Two things fell out of that structure that no sentiment read would produce.

First, the direction of travel on export control is the opposite of the consensus. The common narrative is continual tightening. The record says the January 2026 rule loosened access, moving the H200 to case-by-case review with a tariff attached, and roughly ten Chinese firms including Alibaba, Tencent and ByteDance were cleared to buy. Meanwhile Huawei ships the Ascend 950PR with a CUDA-compatible stack and ByteDance has committed billions in orders. China is not choosing between the two stacks. It is buying the American one while the domestic one matures. That finding cuts against KXCO's own earlier reading of a sanction-proof Chinese stack, and both readings are visible in the map, because a graph that only agrees with itself is not evidence of anything.

Second, compute is leaving the jurisdictional map entirely. On 4 August 2026, on its first earnings call as a public company, SpaceX committed to building its AI compute exclusively on Nvidia's Vera Rubin architecture, and to deploying it in orbit as well as on the ground via a co-designed satellite payload. The policy layer immediately made the consequence legible: export rules, data-localisation requirements and AI regulation are all written for hardware sitting on somebody's soil. Orbital compute sits outside every one of them. That observation only exists because the policy layer was already there to be contradicted.

07 · EconomicsCommoditised compute, proprietary context

Accelerators are a horizontal utility with a spot price. Frontier model access is an API call with a rate card. Neither is a moat, because neither is scarce to anyone with a budget.

A correct, sourced, time-aware map of your own domain is scarce, for reasons that are structural rather than clever:

• It cannot be scraped. The relationships that matter most to an institution, its exposures, counterparties, internal dependencies and obligations, appear in no filing and no news feed.
• It compounds. Every entity added makes every existing entity more useful, because value sits in the edges and edges grow faster than nodes.
• It is expensive in judgement, not compute. The cost is deciding what is true and sourcing it. That is exactly the cost a competitor cannot shortcut by spending more on hardware.
• It appreciates as models improve. Every capability gain in frontier models raises the return on having good structure to point them at. Compute spend depreciates; context spend gets more valuable.

If KXCO tried to win by training a larger model, we would lose, and quickly. The bet is the other way round: that the durable asset in this cycle is the map, not the engine.

08 · The caseWhy your organisation should invest in an ontology

This is the practical section. If you run a fund, a bank, a broker, an exchange, an insurer, a family office or any business with a supply chain, here is the argument in the terms you would actually evaluate it in.

What it replaces

Most organisations already spend heavily on the same job done worse. Research subscriptions that arrive as prose you cannot query. Analysts rebuilding the same dependency map every time somebody leaves. A data room nobody can traverse. A risk report whose conclusions cannot be traced to their inputs. An ontology does not add a line item so much as consolidate several into an asset you own.

The four returns

Institutional knowledge stops walking out of the door. When an analyst leaves, what leaves with them is the mental model, not the spreadsheets. Structured claims persist. This is the difference between a team that knows something and an organisation that does.

Analysis becomes auditable rather than persuasive. Every conclusion decomposes into the claims that produced it, each with a source and a date. When a regulator, an investment committee or a client asks why you believed something, the answer is a path, not a recollection. For regulated institutions this alone frequently carries the business case.

You build the asset once and serve both audiences. The same structure that lets a human traverse a dependency lets an AI agent do it with far less inference and far less room to hallucinate. Organisations deploying agents without a domain ontology are asking a model to guess at facts nobody wrote down, then acting on the guess.

The decay becomes visible. Every organisation's internal knowledge is quietly rotting right now, and nobody can see which parts. Bi-temporal claims make staleness a queryable property instead of an unpleasant surprise.

What it costs, honestly

The expensive part is not software. It is the judgement required to decide what is true, source it, and refuse to fill gaps. A serious ontology is an editorial discipline with a database attached, and any vendor who tells you it is a purely technical project has not built one. Budget for the discipline.

Build or buy

Building in-house is viable if you already have graph engineering capability and, more importantly, the editorial culture to hold the sourcing line under deadline pressure. Most organisations have the first and not the second. The failure mode is not a bad database. It is a graph nobody trusts because at some point somebody filled a gap with an estimate and did not mark it.

09 · Why usWhy KXCO is built to deliver this

Four reasons, each of which you can check rather than take on trust.

1. One of ours is public, so you can audit the method before you buy it

Most vendors show you a demonstration on their data. KXCO publishes a working ontology, free, at kxco.ai/ontology-live, with the whole graph downloadable as one file. You can check any claim against its source, find the gaps, and see the confidence labels on the weak ones. If the method does not survive that, do not buy it. Very few vendors in this category will let you run that test.

Be clear about what that public map is, though. It is a snapshot, it is bounded by public sources, and it is the smallest thing we run. The systems KXCO operates internally and builds for clients are live, ingest continuously, and are far closer to full reality, because they carry the material that never reaches a filing. The public map exists so you can audit the method before trusting the instrument. It is not the instrument.

2. The discipline is demonstrable, because it is visible in what we refused to build

Anyone can claim rigour. It shows up in omissions. The public map does not contain suspected-evasion edges against firms that hold valid export licences, because alleging evasion against a licensed buyer would be false. It does not contain GPU unit counts derived from capital-expenditure dollars, because that arithmetic invents a figure no filing states. It does not model motive, because no source can attest to intent. And it does not assert a contradiction of our own authorship. Each of those was a reasonable-sounding request that did not survive contact with the sourcing standard.

3. The verification stack already exists

An ontology is only as good as your ability to prove what it said and when. KXCO's other capabilities were built for exactly that: post-quantum document signing and data rooms through Nexus, attestation through Sentinel, and a settlement layer and public record in Armature L1, using the NIST-ratified post-quantum standards. Claims can be signed and anchored, so record time is not merely asserted by us. Most graph vendors have no answer here at all.

4. It is built by operators

KXCO's ontology work is led by Shayne Heffernan and John Heffernan, from decades of reading markets rather than from a research lab. The design choices reflect that: two time axes because positions are marked continuously and history matters; confidence labels because a weak claim you can see beats a strong claim you cannot check; and no scorekeeping on short windows, because a twelve-month view marked at three weeks is a position with eleven months left to run, not a result.

10 · LimitsWhat it will not do

Stated plainly, because a vendor who lists no limits is selling something.

It does not tell you what to buy. It identifies market gaps. That distinction is the product, not a hedge. The public map names four: the sovereignty gap, where every participant depends on a stack it does not control; the verification gap, where trust is the scarce asset because demand is partly marking its own homework; upstream chokepoint value, as open weights commoditise the model layer and durable value concentrates in lithography, fabrication, memory and energy; and provenance for defence AI, where models are moving onto classified networks with unaudited supply chains. A gap tells you what is missing from a market. A price target tells you what a stranger thinks a share is worth. Only one of those is structural.

The public map is deliberately incomplete. It covers the spine, not every capillary, and it is bounded by what can be sourced from already-published material. That bound is the point of a public demonstration, and it is also its ceiling.

The public map is not live. It is a periodic snapshot with the refresh date stamped on it. What KXCO runs internally and deploys for clients is live and continuously ingesting, and is materially closer to full reality precisely because it is not restricted to public sources. Anyone judging the method from the public map is judging the smallest version of it.

It will not settle an argument about why someone did something. We record what was said, where and when.

The chat on the public map is the weakest part of it. It answers against the graph rather than the open internet, which is the right design, but the prompt it currently receives includes entities, claims and revelations and does not include the findings or the valuation table. Ask it about either and it will answer from general knowledge and sound confident doing it. That is a real defect on a page whose whole argument is that confident guessing is the problem. We would rather name it than be caught at it. Until it is fixed, use the chat to explore entities and relationships, and read the findings and valuation views directly.

The confidence labels are also honest judgement rather than a calibrated score. 407 claims are marked high, 167 medium and 6 estimate, assigned by the people who sourced them. The label tells you how much weight the sourcing bears, not a probability.

And it will not replace judgement. It removes the excuse for bad judgement by making the inputs checkable, which is a different and more useful thing.

11 · QuestionsQuestions readers ask

What is an ontology in AI?
An ontology is an explicit specification of a conceptualisation. In practice it answers three questions: what things exist, how they are related, and what that means. Unlike a document store or a vector index, an ontology encodes typed relationships between named entities, so a machine can traverse the structure rather than guess at it from text.

Why does context beat more compute?
Because most institutional questions are structural, not linguistic. Asking which of your positions route through a single lithography vendor is a graph traversal, not a language problem. A larger model does not make an unwritten dependency appear; a mapped dependency answers the question with almost no inference at all.

What is a bi-temporal claim?
A claim that carries two independent time axes: valid time, meaning when it held in the world, and record time, meaning when it was asserted. Collapsing those into one date is how research quietly becomes unfalsifiable, because you can no longer tell what was believed when.

How is this different from RAG?
Retrieval-augmented generation fetches text and hopes the model reads it correctly. An ontology hands the model resolved entities and typed relationships, each with a source and a date. RAG retrieves passages; an ontology supplies structure. They compose well, and the ontology is the part that can be audited.

Can an LLM just build the ontology itself?
It can draft one, and it will fabricate. The value of a claim graph is that every edge has a source that supports it, so any model-extracted claim must pass a human verification gate before entering the graph. Without that gate you have a confident-sounding graph, which is worse than no graph.

Why should an organisation invest in an ontology?
Four reasons: institutional knowledge stops leaving when people do; analysis becomes auditable instead of persuasive; the same structure serves humans and AI agents so you build the asset once; and the compounding is real, because every entity added makes every existing entity more useful.

Is the KXCO ontology public?
One of them is, and it is the smallest. The AI-sector map at kxco.ai/ontology-live is free and public, with a machine-readable copy at /ontology-live/data.json, because the AI supply chain can be sourced from public material and therefore demonstrates the method without exposing anyone's data. It is a periodic snapshot and deliberately incomplete. What KXCO runs internally and builds for clients is live rather than snapshotted, and materially closer to full reality, because it is not limited to what has already been published.

Does the ontology make discoveries on its own?
No, and claiming otherwise would misdescribe it. An ontology is an instrument rather than an oracle. It renders structure in a form human judgement can operate on, and the discovery is the person's. The most important finding in KXCO's public map, that the sector converges on a single lithography vendor, was made by a person looking at the display; every underlying fact had been public for years. Without the rendering, nobody goes looking. That is the contribution: making an insight reachable, not automating it.

Does an ontology tell you what to buy?
No, and that is not its job. It identifies market gaps: the dependency with no second source, the layer commoditising while value migrates upstream, the thing everybody needs and nobody can verify. The public map names four such gaps. A gap tells you what is missing from a market, which is more actionable than an opinion on what a share is worth.

What will an ontology not do?
It will not settle a question about motive, because no source can attest to intent, and it will not give you a verdict on a short window. It also will not invent a number to fill a gap; a visible unknown is worth more than a plausible estimate.

12 · SourcesReferences

1. Bureau of Industry and Security, Department of Commerce revises license review policy for semiconductors exported to China.
2. CNBC, US clears H200 chip sales to 10 China firms, May 2026.
3. CNBC, Chip stocks shed more than $1 trillion as selloff hits companies powering AI boom, 29 July 2026.
4. Uncover Alpha, Amazon, Google, Microsoft, Meta Q2 earnings, on 2026 hyperscaler capital expenditure.
5. IEEE Spectrum, China's AI chip race, on the Huawei Ascend line.
6. Inside Privacy, China releases new labeling requirements for AI-generated content.
7. CNBC, SpaceX Q2 2026 earnings, on the exclusive Nvidia commitment and orbital compute.
8. CNBC, Nvidia and OpenAI in talks for up to $250 billion AI backstop, July 2026.
9. CoinDesk, Nvidia forms 37-member AI security alliance, and the Cloud Security Alliance research note.
10. Electrek, Musk pushes unsupervised FSD for consumer Teslas again, April 2026, and Musk says unsupervised FSD will be widespread by year-end, May 2026.
11. Business Wire, Palantir reports Q2 2026 results.
12. Defense One, The Pentagon's bet on autonomous warfare, on the dissolution of Replicator into the Defense Autonomous Warfare Group.
13. T. R. Gruber, A translation approach to portable ontology specifications, Knowledge Acquisition 5(2), 1993, for the definition used here.
14. A. Kollegger, Neo4j, reflections on the AI Engineer World's Fair 2026, on controlled language and agent memory; and F. Coyle, Why Agentic Systems Need Ontologies.
15. KXCO Ontology Live, machine-readable graph, and the guide to reading it.

13 · NextGet more information

The public map is the shortest route to judging the method. Open kxco.ai/ontology-live, pick any claim, and follow it to its source. If a claim's source does not support it, that is a real defect and we would rather hear it from you than not.

If you want to discuss an ontology on your own data, whether that is a portfolio, a counterparty network, a supply chain or a regulatory obligation, the fastest way in is the contact page. Tell us what question you cannot currently answer and we will tell you honestly whether this is the right tool for it.

Nothing in this article is investment advice. Valuation and market figures are third-party data, collected and dated, and are not KXCO forecasts. Entity and claim counts are as at 5 August 2026 and are checkable against the machine-readable graph.