# THK Digital

> A guide for AI agents and crawlers reading https://thkdigital.com. It lists every page on the site and what is on it. It is generated from the pages themselves each time the site is built, so it matches what is published.

- Site map, with the same contents in XML: https://thkdigital.com/sitemap.xml
- Crawling rules: https://thkdigital.com/robots.txt

## About this site

THK Digital Ltd is Tom Knight's practice, registered in England and Wales and based in Warwickshire, UK. The site publishes his working papers on AI architecture and sets out how to work with him.

## Reading notes

- Every page is static HTML; no sign-in, no JavaScript needed to read it.
- Where a page marks a claim "Flagged · Unproven", the author is saying it is not yet proven. Report it that way.
- Each paper's contents entries below link straight to that part of the page.

## Pages

### Tom Knight · AI strategy advisor · THK Digital

- URL: https://thkdigital.com/
- Summary: Tom Knight, AI strategy advisor: wins today, a route to AI baked into the business, and no dependence on one model provider. Working papers in full.

On this page:

- FIG 00 · Advisor · Warwickshire · UK: What your company needs from AI. What AI needs from your company.
  - Link: Work with me: https://thkdigital.com/work-with-me/
- FIG 01 · The working papers · Published in full: The thinking, written in public.
  - Link: Read P.01: https://thkdigital.com/research/infrastructure-first-ai/
  - Link: Read P.02: https://thkdigital.com/research/compute-substrate/
  - Link: Read P.03: https://thkdigital.com/research/memory-wall/
- FIG 02 · Bolt on, bake in: Wins today. Baked in for good.
- FIG 03 · No single point of failure: One model provider is one point of failure.
- FIG 04 · Where the line goes: Most of your operation is ordinary work.
- FIG 05 · Field record · Keystone Integrations: I learned it the expensive way.
  - Link: The full history: https://thkdigital.com/about/
- FIG 06 · Provenance: You know who wrote it.
  - Link: start here: https://thkdigital.com/investors/
  - Link: About me: https://thkdigital.com/about/

Flagged claim (Flagged · Unproven): Baked in at scale is still a bet, and I mark it as one: I have not delivered it end to end yet, and I will not claim it until I have. The papers show where the evidence stands.

### About Tom Knight · THK Digital

- URL: https://thkdigital.com/about/
- Summary: Tom Knight, AI strategy advisor and researcher. Founder of THK Digital; before that, co-founder of Keystone Integrations and an engineer at PDSL.

On this page:

- FIG 00 · About · Stratford-upon-Avon · Warwickshire: Tom Knight.
  - Link: For investors: https://thkdigital.com/investors/
- FIG 01 · Work: The history.
- FIG 02 · Education: Education.
- FIG 03 · Writing and speaking: Writing and speaking.
  - Link: P.01 · 2026-05 · Why the Model Should Be the Smallest Part of Your Stack: https://thkdigital.com/research/infrastructure-first-ai/
  - Link: P.02 · 2026-06 · Everyone is Hammering Screws: https://thkdigital.com/research/compute-substrate/
  - Link: P.03 · 2026-11 · The Memory Wall: https://thkdigital.com/research/memory-wall/
- FIG 04 · Contact: Get in touch.
  - Link: LinkedIn: https://www.linkedin.com/in/tom-knight-93178126b/

### Investors · THK Digital

- URL: https://thkdigital.com/investors/
- Summary: Tom Knight is building a company, in stealth, alongside his advisory work and research. What he looks for in investors, and how to get in touch.

On this page:

- FIG 00 · Building: I am building a company, in stealth.
  - Link: Read the papers: https://thkdigital.com/research/
- FIG 01 · What I am looking for: More than money.
- FIG 02 · Tell me about you: Who you are, and what you bring.

### Research · THK Digital

- URL: https://thkdigital.com/research/
- Summary: Tom Knight's working papers on AI as infrastructure, published in full, with what is proven and what is not marked in the text.

On this page:

- FIG 00 · Research index: Notes from the build.
  - Link: P.01 · 2026-05 · Why the Model Should Be the Smallest Part of Your Stack. · Thesis · ~35 min read: https://thkdigital.com/research/infrastructure-first-ai/
  - Link: P.02 · 2026-06 · Everyone is Hammering Screws. · Thesis · ~45 min read: https://thkdigital.com/research/compute-substrate/
  - Link: P.03 · 2026-11 · The Memory Wall. · Thesis · ~80 min read: https://thkdigital.com/research/memory-wall/

### Work with me · THK Digital

- URL: https://thkdigital.com/work-with-me/
- Summary: AI strategy from Tom Knight: where AI earns its cost in your business, what it needs from you, and a plan your own people carry. Priced on the outcome.

On this page:

- FIG 00 · Work with me · C-suite AI strategy: AI is expensive. The talent is limited. Where does your money go first?
- FIG 01 · People first: The technology is incredible. Your people matter more.
- FIG 02 · The plan: Bolt on now. Bake in over time.
- FIG 03 · How it works: You are paying for the outcome, not my hours.
- FIG 04 · Why take my word for it: You do not have to. The reasoning is published.
  - Link: Read P.01: https://thkdigital.com/research/infrastructure-first-ai/
  - Link: Read P.02: https://thkdigital.com/research/compute-substrate/
  - Link: Read P.03: https://thkdigital.com/research/memory-wall/
- FIG 05 · Get in touch: Start with a conversation.

## Working papers

### Working Paper 01: Why the Model Should Be the Smallest Part of Your Stack · THK Digital

- URL: https://thkdigital.com/research/infrastructure-first-ai/
- Author: Tom Knight
- Published: 12 May 2026
- Reading time: ~35 min read
- Figures: 1
- Cite as: Tom Knight, "Why the Model Should Be the Smallest Part of Your Stack · THK Digital", THK Digital Working Paper 01, 12 May 2026, https://thkdigital.com/research/infrastructure-first-ai/

Summary: The dominant narrative around agentic AI in 2026 is that the models are not yet good enough. The standard reading is wrong. The technology has not been over-sold; it has been under-engineered. The models are not the problem. The architectures wrapped around them are.

Contents:

- 1. Thesis: https://thkdigital.com/research/infrastructure-first-ai/#thesis
- 2. The Bolted-On Failure Mode: https://thkdigital.com/research/infrastructure-first-ai/#bolted-on-failure-mode
- 3. The Five Primitives of Infrastructure-First AI: https://thkdigital.com/research/infrastructure-first-ai/#five-primitives
- 4. Why This Architecture Enables Small Language Models: https://thkdigital.com/research/infrastructure-first-ai/#slm-enabling-architecture
- 5. The Strategic Case for Infrastructure Independence: https://thkdigital.com/research/infrastructure-first-ai/#strategic-case
- 6. Case Study: A Production Scheduling Agent: https://thkdigital.com/research/infrastructure-first-ai/#case-study
- 7. Future Work: From Architecture to Ownership: https://thkdigital.com/research/infrastructure-first-ai/#future-work
- 8. What Comes Next: https://thkdigital.com/research/infrastructure-first-ai/#what-comes-next
- References: https://thkdigital.com/research/infrastructure-first-ai/#references

### Working Paper 02: Everyone is Hammering Screws · THK Digital

- URL: https://thkdigital.com/research/compute-substrate/
- Author: Tom Knight
- Published: 23 Jun 2026
- Reading time: ~45 min read
- Cite as: Tom Knight, "Everyone is Hammering Screws · THK Digital", THK Digital Working Paper 02, 23 Jun 2026, https://thkdigital.com/research/compute-substrate/

Summary: This paper makes a single load-bearing claim: where a computation physically runs fixes what it costs and how long it takes, and no amount of clever software above that layer can unfix it. Infrastructure-first architecture is the discipline of routing each computation to the silicon that handles it cheapest and fastest, and reserving the most expensive silicon in the building for the one job that actually requires it. The industry's bill is not the price of intelligence. It is the price of swinging the hammer at every job.

Contents:

- 1. Thesis: https://thkdigital.com/research/compute-substrate/#thesis
- 2. The Compute Substrate: https://thkdigital.com/research/compute-substrate/#compute-substrate
- 3. The Retrieval Substrate: https://thkdigital.com/research/compute-substrate/#retrieval-substrate
- 4. The Cost of the Hammer: https://thkdigital.com/research/compute-substrate/#cost-of-the-hammer
- 5. Efficiency Is a Choice: https://thkdigital.com/research/compute-substrate/#efficiency-is-a-choice
- 6. Example Scenario: https://thkdigital.com/research/compute-substrate/#example-scenario
- 7. What Comes Next: https://thkdigital.com/research/compute-substrate/#what-comes-next
- References: https://thkdigital.com/research/compute-substrate/#references

### Working Paper 03: The Memory Wall · THK Digital

- URL: https://thkdigital.com/research/memory-wall/
- Author: Tom Knight
- Published: 14 Nov 2026
- Reading time: ~80 min read
- Figures: 21
- Download the PDF: https://thkdigital.com/research/memory-wall/thesis3-memory-wall.pdf (48 pages · 3.2 MB)
- Cite as: Tom Knight, "The Memory Wall · THK Digital", THK Digital Working Paper 03, 14 Nov 2026, https://thkdigital.com/research/memory-wall/

Summary: Get the result on a box you own, without data-centre memory. Claim the job, not the model.

Contents:

- Executive summary: https://thkdigital.com/research/memory-wall/#executive-summary
- How to read this paper: https://thkdigital.com/research/memory-wall/#how-to-read
- Physics and scale: a quick reminder: https://thkdigital.com/research/memory-wall/#physics-and-scale
- Opening: https://thkdigital.com/research/memory-wall/#opening
- 1. Thesis: https://thkdigital.com/research/memory-wall/#thesis
- **Act one: what agents in production need**: https://thkdigital.com/research/memory-wall/#act-one
- 2. Four kinds of agent, four different bills: https://thkdigital.com/research/memory-wall/#four-kinds-of-agent
- 3. “Reactive” means two things: https://thkdigital.com/research/memory-wall/#reactive
- 4. What agents hold: the model, plus the memory of every conversation: https://thkdigital.com/research/memory-wall/#what-agents-hold
- 5. Why memory, not maths: the fetch: https://thkdigital.com/research/memory-wall/#the-fetch
- 6. How memory was moved closer, and where that ends: https://thkdigital.com/research/memory-wall/#memory-moved-closer
- 7. Why it fails in production: the wait before the first word: https://thkdigital.com/research/memory-wall/#wait-before-first-word
- **Act two: eight ways to cut the memory, judged from your own box**: https://thkdigital.com/research/memory-wall/#act-two
- 8. Eight methods: https://thkdigital.com/research/memory-wall/#eight-methods
  - 8.1 Mixture of experts: fire fewer weights per token: https://thkdigital.com/research/memory-wall/#mixture-of-experts
  - 8.2 Quantisation: fewer bits per weight: https://thkdigital.com/research/memory-wall/#quantisation
  - 8.3 Offloading to system RAM: https://thkdigital.com/research/memory-wall/#ram-offload
  - 8.4 Streaming from the SSD: https://thkdigital.com/research/memory-wall/#ssd-streaming
  - 8.5 Batching: https://thkdigital.com/research/memory-wall/#batching
  - 8.6 More silicon: https://thkdigital.com/research/memory-wall/#more-silicon
  - 8.7 Shrinking the KV cache: https://thkdigital.com/research/memory-wall/#shrinking-kv-cache
  - 8.8 Small language models: need less to begin with: https://thkdigital.com/research/memory-wall/#small-models
- 9. Where every method breaks: the same place: https://thkdigital.com/research/memory-wall/#same-place
- **Act three: the alternative**: https://thkdigital.com/research/memory-wall/#act-three
- 10. The shared data centre is a bigger box that everyone shares: https://thkdigital.com/research/memory-wall/#shared-data-centre
- 11. One principle: hold the memory, reshape the work: https://thkdigital.com/research/memory-wall/#one-principle
- 12. Route each agent to the model it needs: https://thkdigital.com/research/memory-wall/#route-each-agent
- 13. Worked example: the frontier model on your box against a model built for the job: https://thkdigital.com/research/memory-wall/#worked-example
- 14. What this does not claim, and what would kill it: https://thkdigital.com/research/memory-wall/#does-not-claim
- 15. What comes next: https://thkdigital.com/research/memory-wall/#what-comes-next
- References: https://thkdigital.com/research/memory-wall/#references-heading
