Skip to the first chapter
Next word → real work
A plain-English tour of the AI industry

From next word
to real work.

AI as a building: what it runs on, who owns each floor, and the fit-out that turns a clever model into completed work.

By Ambrish NarayanR.E.N Reviews15-minute read
2020The tenant signed a five-year …lease
2023Summarise this lease and list the break clausesdone in seconds
2025Prove this olympiad problem, step by stepgold medal
2026Find a blow-up in the Navier-Stokes equationsclaimed, under review
Illustration: an office tower cut open to show AI’s floors, from power and data centres up to people at work
How every model writes: one word at a time

The tenant signed a five-year lease

lease82%
contract8%
deal6%
agreement4%

Illustrative probabilities. The model scores every possible next word, picks one, then repeats.

How far it has come2 min

It started by guessing the next word.

Every model today still works by predicting what comes next. Scale, training on reasoning and the ability to use tools turned that simple trick into systems that do real work.

1.2bnpeople use ChatGPT every weekOpenAI DevDay, 29 Sep 2026
1bnGemini app usersAlphabet Q3 results, Oct 2026
30m+paid Microsoft 365 Copilot seatsMicrosoft FY26 Q4, Jul 2026
22bntokens a minute through Google’s model APIsAlphabet Q2 results, Jul 2026
Illustration for the predict era

Predict

2017-2020

Models learn the patterns of language, and of biology.

2017

The Transformer

Google researchers publish “Attention Is All You Need”. Every major model since is built on it.

2020

GPT-3: next-word prediction at scale

175 billion parameters. Fluent text, with no real reasoning yet.

2020

AlphaFold 2 predicts protein structures

Unveiled at the CASP14 competition in November 2020: a 50-year biology problem effectively solved. Demis Hassabis and John Jumper share the 2024 Nobel Prize in Chemistry.

Illustration for the adoption era

Adopt

2022-2023

AI becomes generally available, and everyone starts using it.

2022

ChatGPT opens to the public

Released free in November 2022, it reached an estimated 100 million users in about two months: the fastest-growing consumer app of its time.

Illustration for the reason era

Reason

2024-2025

Step-by-step thinking unlocks maths and logic.

2024

Reasoning models

Models learn to think before answering, working through problems step by step.

2025

Gold at the International Maths Olympiad

OpenAI and Google DeepMind models reach gold-medal standard on the world’s hardest school maths competition.

Illustration for the act era

Act

2025-2026

AI stops just answering and starts doing.

2025-26

Agents

AI that uses tools, writes and runs code, reads your files and completes multi-step tasks.

Illustration for the discover era

Discover

2026

The first claims of genuinely new results, still being tested.

Sep 2026

ClaimedA Millennium Prize problem

OpenAI announces a blow-up result for the Navier-Stokes equations, one of maths’ $1m problems. About 10,000 AI agents worked for 88 hours; the proof is machine-checked but still under expert review, and credit is disputed.

Sep 2026

In trialAn AI-discovered drug reaches Phase III

Insilico’s rentosertib, for lung fibrosis, with both target and molecule found by AI, doses its first Phase III patient. No AI-discovered drug has full FDA approval yet.

Why it keeps going2 min

Predicting the next word is how it’s trained. It isn’t a ceiling.

Many models still generate one token at a time. Training, reasoning, tools and feedback expand what the complete system can do. You can see it in the test scores.

The tests keep getting harder. The scores keep climbing.

Best published score by any AI model on four well-known tests, by date. Each new test was written because the last one got too easy.

Best published AI score on four well-known tests, by dateMMLU: from 43.9% to 93.3%; GPQA Diamond: from 35.7% to 96%; ARC-AGI-2: from 4% to 95%; Humanity’s Last Exam: from 7.96% to 54.8%2020202120222023202420252026Today0%25%50%75%100%93%MMLU96%GPQA Diamond95%ARC-AGI-255%Humanity’s Last Exam

Each line is a test. Each dot is a new best score. Hover or tap a dot for the model and date.

4% → 95%

ARC-AGI-2, in about 17 months

Puzzles most people can solve but AI found very hard. From its launch in Mar 2025 to Sep 2026.

8% → 55%

Humanity’s Last Exam

Written in Jan 2025 to be the hardest exam yet. The best AI still gets about 45% of it wrong.

Exams ≠ jobs

Test scores, not job scores

A benchmark is an exam with known answers. Real work also needs context, judgement, permissions and accountability. That’s the harness, and it’s where people stay in charge.

Each dot is a new best published score. GPQA Diamond and MMLU scores are from AI companies’ reports and the original papers; Humanity’s Last Exam is from Scale AI’s leaderboard (no tools); ARC-AGI-2 is from ARC Prize’s verified results. Labs test under different conditions, and their own reports often run higher than independent checks, so treat individual points as approximate. No MMLU record after April 2025 could be verified. Scores checked 3 November 2026.

Three forces behind the progress

Compute alone gives steady gains. Three forces push the curve upward.

Illustrative
Illustrative: three forces push AI progress from a straight line into an upward curveCompute alone gives steady gains, a straight line. More resources, better methods and AI helping AI research each push the curve higher, bending it upward. The final part is dashed because compounding self-improvement is still a hypothesis.CapabilityTimeIllustrative shape, not measured data010203Compute alone: steady gainsExponential, if AI-driven research compoundsThe recursive self-improvement question: still a hypothesis
01

Scale the resources

Spending on chips, data and bigger models keeps rising, and models now ‘think’ for longer when they answer.

02

Improve the methods

Better architectures, training techniques, search and verification get more useful capability from the same resources.

03

Use AI in the research

Systems propose, test and refine improvements to code and algorithms, including parts of the machinery used to build AI. If that loop compounds, progress accelerates.

GenerateEvaluateSelectIterate

Evidence for force 03: AlphaEvolve

Google DeepMind reports gains in data-centre operations, chip design and training code from a model-guided, evaluator-tested search loop.

Established: AI-assisted improvement in bounded systems. Still a hypothesis: self-sustaining runaway acceleration.

What it runs on2 min

AI is a building. Read it from the basement up.

Every chat answer sits on a tower of physical infrastructure. Each floor has its own economics, and its own winners. Scroll, and watch it go up.

B1plant room
Energy & gridBasement
Gthe shell
Data centresGround
L1the engines
ChipsLevel 1
L2the landlords
Cloud providersLevel 2
L3the brains
ModelsLevel 3
PHthe tenants
Apps & agentsPenthouse
B1Basement · plant room

Energy & grid

The real bottleneck. Grid connections, generation and cooling now decide where AI gets built.

2×data-centre electricity demand by 2030; up 17% in 2025 aloneInternational Energy Agency, Apr 2026
GGround · the shell

Data centres

Powered, cooled shells for GPUs. A single AI campus can now draw as much power as a city.

L1Level 1 · the engines

Chips

NVIDIA GPUs dominate, with Google TPUs and in-house chips from Amazon and Microsoft. TSMC makes most of them.

$96.2bnNVIDIA revenue in one quarter, up 106%NVIDIA Q2 FY27, 26 Aug 2026
L2Level 2 · the landlords

Cloud providers

Often called the hyperscalers. Amazon (AWS), Microsoft (Azure), Google Cloud and Oracle fill data centres with chips and rent that computing power, by the hour, to the model makers and businesses on the floors above. SpaceX, which now owns xAI, has joined them.

~$800bnBig Tech capex in 2026Business Insider tally of guidance, 30 Oct 2026
What the landlords are spending on AI in 2026
Amazon~$220bn
Alphabet$195-205bn
Microsoftcalendar year~$170bn

Alphabet’s 2027 capex is signalled above $300bn. Bars scaled to Amazon. Sources: Business Insider (30 Oct 2026); Alphabet Q3 2026 results.

L3Level 3 · the brains

Models

GPT-6, Claude, Gemini, Grok, DeepSeek, Qwen, GLM. Frontier labs, plus open models anyone can run.

11,000+models in Microsoft’s catalogueMicrosoft, Jul 2026
PHPenthouse · the tenants

Apps & agents

Where people meet AI: ChatGPT, Gemini, Copilot, Claude, Grok, coding agents like Cursor and Codex. Between the brains and the tenants sits the fit-out. We’ll come back to that.

1.2bnweekly ChatGPT usersOpenAI, Sep 2026
Who owns which floor2.5 min

Seven players, seven different bets.

The people steering AI’s biggest companies are each betting on different floors of the building. Here is what they’re worth, what they own, and what they’re betting on.

What they’re worth (US$ trillion)
NVIDIAlisted4.94
Microsoftlisted3.88
Alphabetlisted3.42
SpaceX + xAIlisted Jun 20261.98
Teslalisted1.58
Anthropicprivate round0.97
OpenAIprivate round0.85
market cap, late Oct 2026last private valuationOpenAI in talks at $1.25tn (Oct 2026)

For scale: Apple overtook NVIDIA as the world’s most valuable company in October 2026 as investors questioned AI returns. Listed market caps and private funding rounds are not like-for-like, and both move daily.

Jensen Huang
Jensen HuangCo-founder, NVIDIA · CEO until Nov 2026
NVIDIA logo
$4.94tnmarket cap, late Oct 2026

The bet: Sell the engines to everyone. Whoever wins the model race, they buy NVIDIA.

Jensen Huang
Floors

Chips, plus full systems and software (CUDA)

Latest

$96.2bn quarterly revenue; Vera Rubin chips in full production

Latest move · 12 Nov 2026

Stepped down as CEO after 33 years. CFO Colette Kress is interim CEO while the board runs a search. The open question: whether the relationships he held stay with the company. Q3 results are due on 17 November.

Leadership change
Sam Altman
Sam AltmanCEO, OpenAI
$852bnlast round (early 2026); in talks at $1.25tn, Oct 2026

The bet: The biggest consumer AI brand, scaled to general intelligence.

Sam Altman
Floors

Models and apps, moving into data centres (Stargate)

Latest

GPT-6 (Oct 2026); 1.2bn weekly users

Latest move · 13 Nov 2026

Launched ChatGPT Wallet, letting agents make purchases for US Pro users, days after buying Astral (the makers of uv and Ruff) to strengthen Codex.

Rolling out, US Pro first
Dario Amodei
Dario AmodeiCEO, Anthropic
$965bnSeries H, May 2026 ($65bn raised)

The bet: Win business and coding work, with safety research as the brand.

Dario Amodei
Floors

Models and apps (Claude, Claude Code)

Latest

$47bn revenue run-rate (May 2026); preparing to list

Latest move · 10 Nov 2026

Launched Claude Meetings, which joins video calls and turns the discussion into plans and tasks, a day after releasing Claude Fable 5.2.

Meetings in beta for Team and Enterprise
Elon Musk
Elon MuskSpaceX (incl. xAI, X, Grok) · Tesla
$1.98tnSpaceX, 16 Oct 2026; Tesla a further $1.58tn

The bet: Own the whole vertical, from power and rockets to robots.

Elon Musk
Floors

Every floor: energy (Tesla), data centres (Colossus), models (Grok), apps (X, Cursor)

Latest

xAI merged into SpaceX (Feb 2026); SpaceX listed in June. Tesla remains a separate company.

Latest move · 11 Nov 2026

Released Grok Bot 1.0 for Slack and Microsoft Teams. Musk says Grok 5 is two or three months away.

Reported release; Grok 5 not yet out
Sundar Pichai
Sundar PichaiCEO, Alphabet (Google)
$3.42tnmarket cap, late Oct 2026

The bet: The only player that makes its own chips, models and apps at scale.

Sundar Pichai
Floors

Every floor in-house: TPU chips, cloud, Gemini, apps, DeepMind science

Latest

Gemini app at 1bn users; Cloud revenue up 103% in Q3

Latest move · 10 Nov 2026

Launched Gemini for Healthcare, an agent that reads patient records from Epic and Oracle Health systems.

Preview for Gemini Enterprise customers
ChinaDeepSeek · Alibaba Qwen · Moonshot Kimi · Zhipu GLM
Openmodel weights released free to download

The bet: Good-enough models at a fraction of the cost, open to all.

China
Floors

Models and apps; Alibaba Cloud; Huawei chips

Latest

DeepSeek V4 (Apr 2026). NVIDIA now assumes no China data-centre sales

Latest move · 16 Nov 2026

Alibaba launched Qwen4-Max, its new flagship, without releasing the model weights. Tencent and Moonshot have limited cyber-attack abilities in their latest models.

Launched; closed weights
Satya Nadella
Satya NadellaCEO, Microsoft
Microsoft logo
$3.88tnmarket cap, late Oct 2026

The bet: Own the enterprise and stay neutral on models. Use OpenAI, Anthropic, xAI or Microsoft’s own, inside the tools people already use.

Satya Nadella
Floors

Data centres, cloud (Azure), own models (MAI), apps (Microsoft 365 Copilot)

Latest

Azure up 48% in the September quarter; stakes in both OpenAI and Anthropic

Latest move · 17 Nov 2026

Opens Ignite 2026 in San Francisco, three weeks after previewing the Copilot harness in Foundry Agent Service.

Conference, 17-20 Nov

Click a card to see its floors and latest move. Valuations are as at late October 2026; latest moves as at 17 November 2026. Jensen Huang stepped down as NVIDIA’s CEO on 12 November 2026; the bet he set is still the company’s.

Who owns which floor
PlayerEnergyData centresChipsCloudModelsApps
NVIDIA
OpenAI
Anthropic
Musk companies
Alphabet
China
Microsoft
owns itpartly, or through partnersnot a focusTesla is a separate listed company from SpaceX.
Who is paying for the building2 min

The same dollars keep going round. Flywheel, or loop?

A growing share of the building is funded by its own suppliers. It’s called circular financing: the companies selling the chips and the cloud are also funding the customers who buy them.

How the circle works
The supplierChip makerInvests in an AI lab, or guarantees the debt behind its data centre.
→cash and guarantees
The customerAI labSigns multi-year contracts for computing power it can’t yet pay for from revenue.
→compute contracts
The landlordCloud providerBorrows against those contracts to build, then orders the chips.
↩ Chip orders land back with the supplier. Each step is a real transaction. The question is how much of the demand would exist without the supplier’s own money.
~$1tnin interlinked AI deals, about two-thirds of it borrowedIMF managing director, CNBC interview, Nov 2026
$300bnof computing power Oracle agreed to supply OpenAI over five yearsAnnounced 2025, via Global Finance
$105bnof financing NVIDIA agreed to guarantee for a data centre leased by OpenAIAug 2026, via Reuters
$30bnAnthropic’s commitment to Microsoft’s cloud and NVIDIA’s chips, after the two invested about $15bn in itGlobal Finance, updated Aug 2026
12Nov 2026

The circle just got a stress test. Jensen Huang stepped down as NVIDIA’s CEO, and NVIDIA is the largest financier in the loop. Reuters’ read: its $500bn financing platform is unlikely to disappear under interim CEO Colette Kress, but new deals may come slower and on tougher terms. Days later the European Central Bank’s vice-president warned that a rethink of AI expectations could trigger an abrupt repricing in markets.

My read on the risk: much of this web ran on Huang’s personal relationships. Customers quietly building or buying alternatives to NVIDIA’s chips had reason to be wary while he held the top seat. With him gone, some may feel freer to move, and the deals that tie the loop together get harder to hold in place. But the devil is in the detail: NVIDIA’s chips are simply next level, and demand for them will stay high. My hope? Jensen does an Arnie: “I’ll be back.”

Developing
The flywheel case

Real demand, funded early

Suppliers have always financed good customers. The deals are largely made in the open, usage keeps climbing, and the chips are running flat out rather than sitting in warehouses.

The loop case

Revenue you paid for yourself

When a supplier funds its own customer, growth can flatter everyone’s numbers at once. Much of it is debt, and critics argue chips wear out faster than the five to six years they’re depreciated over.

If you’re buying AI

Useful either way

The tools work whether or not the financing unwinds. What could change is price and which suppliers are standing, so keep the ability to switch models. That’s where the harness comes in.

Deal sizes are as announced and often conditional; several are commitments over many years, not cash paid. General information, not financial advice.

The fit-out3 min

A model can reason. A harness makes work possible.

The software around the model supplies context, connects actions and keeps the task moving. Anyone who has fitted out an office already knows this idea.

Illustration: the same office floor fully fitted out around the glowing model, with plans, shelves, tools, a work loop and security gate Illustration: a bare shell-and-core office floor with only the glowing model in the middle Shell & corea brilliant model, on its own Fitted outmodel + harness
The property version

A model on its own is shell and core.

Impressive, valuable, and nobody can work in it yet. It doesn’t know your priorities, your systems or what it’s allowed to do.

The harness is the fit-out: purpose, context, memory, tools, execution and controls. That’s what turns a floor into a workplace, and a clever answer into a completed task.

A change you can feel

We used to assemble the agent. Now we open a workspace.

1

OpenClaw made the machinery visible.

Memory files, tool connections and configuration made an agent’s working environment something you could see and touch.

2

Codex and Claude package more of it.

Desktop workspaces bring files, tools, ongoing tasks and computer use into one place.

3

The interface simplifies. The system remains.

Less setup for the person using it; more responsibility for the platform providing it. The harness is becoming part of the product.

Examples of different approaches, not a claim that one invented the category.

Agent workspace
  • 01Define the instructions
  • 02Connect the tools
  • 03Configure memory
  • 04Schedule the work
  • 05Set permissions
  • 06Check the outcome
Running across files, apps and steps
Six layers around the model
CONTROLSEXECUTIONTOOLSMEMORYCONTEXTPURPOSEMODELthe intelligence
Monthly board updateDraft
  • Approved two-page structure: facts, assumptions, open questions
  • Sources: last month’s figures, project tracker, relevant policy (cited)
  • Scope, finished sections and open questions carried forward
  • Records read, comparison calculated, draft document created
  • Missing source flagged, failed calculation corrected, no duplicate drafts
  • Waiting for the right person to approve before it is shared
Layer 01 · Purpose

Purpose

Give the work a clear purpose.

Set the task, priorities and reusable method. The model needs to know what a good result looks like.

Example: monthly board update“Prepare a two-page board update using the approved template. Separate facts, assumptions and open questions.”
Layer 02 · Context

Context

Bring the right information into view.

Supply relevant records and retrieve permitted documents. Useful context beats an enormous pile of unrelated material.

Example: monthly board updateUse last month’s figures, the project tracker and the relevant policy. Cite the information used.
Layer 03 · Memory

Memory

Keep continuity between steps.

Preserve decisions, preferences and work in progress so the next session can continue. Stored information still needs a source and a shelf life.

Example: monthly board updateKeep the update’s scope, completed sections and outstanding questions. Re-check figures that can change.
Layer 04 · Tools

Tools

Turn an answer into an action.

Connect files, applications, APIs or computer interaction. Tool protocols such as MCP standardise connections; access must still be configured.

Example: monthly board updateRead the permitted records, calculate the comparison and create a draft document for review.
Layer 05 · Execution

Execution

Coordinate the work from start to finish.

Run the model-tool loop, inspect results, handle errors and resume after interruptions. Stop when the task needs human judgement.

Example: monthly board updateIf a source is missing, flag it. If a calculation fails, correct it. Avoid creating the same update twice.
Layer 06 · Controls

Controls

Define what can happen, and what gets checked.

Apply permissions, approval points, evaluations and logs around actions. A capable model does not remove the need for accountability.

Example: monthly board updateDraft the update automatically. Require the right person to approve it before it is sent.
UnderstandChoose a stepUse a toolInspect the resultContinue or askTHE HARNESSRUNS THE LOOP
More than a chat window

The harness runs the loop, and asks a different question.

Model questionCan it work out what to do?
Harness questionCan it do it here, correctly, with permission?
Bringing it into the enterprise2 min

Copilot is the front door. Foundry builds what works behind it.

Satya’s bet is that companies will run AI inside the tools they already use, and choose models the way they choose suppliers.

28Oct 2026

Microsoft introduced the Copilot harness in Foundry Agent Service: the managed loop, sandbox, tools and reusable skills behind Copilot, opened up for customer-built agents.

Public preview
35models

Foundry’s Model Router now picks between 35 models per request. October added GPT-6 Astra, Claude Fable 5.1, Opus 5 and Sonnet 5, Grok 4.6, GLM-5.3 and Kimi-K3.

Model choice, in practice
GAOct 2026

Now generally available: long-running resilience for hosted agents, browser automation, and a Decisions API that returns a yes/no, a choice or a score instead of prose.

Generally available
17Nov 2026

Ignite 2026 opens in San Francisco. The session catalogue leans on three layers: Copilot, Microsoft IQ and Agent 365, with governing and securing agents as the running theme.

Conference, 17-20 Nov
The bet, in one line

Own the enterprise and stay neutral on models: OpenAI, Anthropic, xAI or Microsoft’s own, inside the tools people already use.

Illustration: an employee greeted at the front door by an AI assistant, with organised systems behind a secure glass wall
The experience

Microsoft 365 Copilot · Teams

A familiar place for employees to ask, review and act.

The agent

Copilot Studio · Foundry

Define the task, connect the tools and operate the workflow.

The intelligence

Model choice

Select and evaluate suitable models for each task.

OpenAIAnthropicMicrosoftothers
The connection

Enterprise data and systems

Ground work in permitted information and approved actions.

Still the organisation’s responsibility, across every layer
  • Identity and access
  • Data boundaries
  • Human approval
  • Evaluation and testing
  • Logs and traceability
  • Named process owners

A smarter model expands what’s possible. A well-designed harness turns it into repeatable work.

01 · Where the work already is

AI goes to the data

Documents, data and identities already live in Microsoft 365, SharePoint and Teams. The agent meets people there instead of in another app.

02 · Any model

Choice, not lock-in

OpenAI, Anthropic and other models side by side. Pick the right one for cost, sensitivity and speed, and swap it as the models improve.

03 · Governed by design

Controls you already run

Existing identity, permissions and data-governance controls apply to every agent from day one, rather than being bolted on later.

Conceptual enterprise architecture: not a statement that every Copilot product has identical capabilities, or that every control is configured by default. Previews are a direction, not a finished rollout. Release status as listed by Microsoft on 5 November 2026.

08 · Over to you0.5 min

The model is getting smarter. The job now is to make that intelligence useful.

Give it the right context

Connect the task to the information and process that matter.

Give it the right reach

Make useful actions possible, within clear permissions.

Keep the work accountable

Evaluate outcomes, record actions and keep people in charge.

Three things to remember

  1. AI went from guessing the next word to doing real work in under ten years, and it keeps passing harder tests.
  2. It runs on a physical building of power, data centres and chips, and who pays for that building matters as much as who builds it.
  3. The value is in the fit-out: the harness of context, tools, memory and controls. That’s Microsoft’s bet, and it’s where most organisations will do their own building.
Reference desk

Where the numbers come from

Public facts checked in early November 2026; later items carry their own dates. Market capitalisations and private valuations change daily; all figures are as reported at the dates shown, and private valuations are post-money from the latest priced round. Announced plans and previews are labelled separately from available products. Interpretation is the author’s own.

Usage, money and markets

Capability and research

Harnesses, companies and Microsoft

Benchmark scores (chapter 02)

Image credits. Portraits: Sam Altman and Dario Amodei, TechCrunch, CC BY 2.0 · Elon Musk, Steve Jurvetson, CC BY 2.0 · Sundar Pichai, Lukasz Kobus / European Commission, CC BY 4.0 · Jensen Huang, Peter DaSilva, CC BY 4.0 · Satya Nadella, LeWeb official photos, CC BY 2.0. All cropped for display. Company logos identify the organisations discussed and do not imply endorsement. Illustrations were generated for this presentation with AI.