//PARTNERSHIP · InAI Capital

Build AI.
Fund AI.
Decode AI.

We invest in frontier AI, semiconductors and robotics out of Silicon Valley, publish the industrial research global banks read, and run a quantitative engine that rebuilds every trading day.

Silicon Valley · since 2026 Venture · research · quant A-share + US technology Quant data as-of —
Start a conversation

15 slides, 4 live demos. Use or scroll.

//FOUNDER

Built the systems before backing them.

Yumei Dou — a decade inside the AI industry before moving to the other side of the table. Co-founder of early AI unicorns; deep-learning systems in production well before the LLM era.

01 · OPERATOR

AI unicorn co-founder

Algorithms, product, growth — shipped, not modelled.

02 · INVESTOR

Investing in technology since 2019

AI, semiconductors, robotics; energy and space frontier.

03 · ADVISOR

Institutional advisor

Global investment banks and top-tier VC and PE firms.

THE CONVICTION UNDER ALL FOUR LINES

AI is rewriting the cost curve of every industry it touches — and the people best placed to see it are the ones who have actually shipped the systems.

Full biography →

SELECTED ADVISORY ENGAGEMENTS

Goldman Sachs UBS HSBC Standard Chartered SoftBank GLG Guidepoint

//WHAT WE DO

Four lines. One thesis.

AI is rewriting the cost curve of every industry it touches. Each line reads that from a different distance — and serves a different counterparty.

A

VENTURE

Frontier
equity

Seed to growth in AI silicon, robotics and frontier labs.

serves Startups & labs

B

RESEARCH

Industrial
deep research

Architecture, supply chain, and the Bit‑Watt framework.

serves Banks & institutions

C

QUANT DATA

Daily
signal engine

5,203 A-share stocks plus US technology, rebuilt every close.

serves Investors

D

ENTERPRISE

Licensing &
white-label

The whole engine, running inside your infrastructure.

serves Funds & brokerages

//WHY IT COMPOUNDS

Each line pays for the next one’s edge.

Every link hands over an input the receiving line could not have bought. Three of them are a specific artifact changing hands; the fourth is a relationship — and we mark it as such.

The operating loop
what actually changes hands
Solid arrows hand over an artifact. The dashed one hands over standing.
A · VENTURE Frontier equity Seed to growth, Silicon Valley B · RESEARCH Deep research Architecture and supply chain C · QUANT DATA Daily signal engine Rebuilt after every close D · ENTERPRISE Licensing & white-label Your infrastructure, our engine portfolio access — roadmaps a year before they ship the sector taxonomy 62 sectors, 525 stocks — one file the daily series every claim is checkable against it the five parts, packaged as API, licence or white-label institutional relationships connection & trust

The taxonomy is a literal file. The research line maintains it and the quant engine aggregates by it, so changing the sector map moves both at once. The daily series run the other way: a claim is either visible in them or it is not.

//THE SPINE

Bit → Watt → Execution.

One framework sits under all four lines, and everything after this slide rests on it. The test is narrow on purpose: does this company use a signal to control power? Here is what it produces.

Process node vs track tier — semiconductor atlas
Bit‑Watt atlas · —
The atlas names 66 companies (30 A-share, 36 US). Plotted here are the 37 that are semiconductors with a process node — equipment makers, power and battery names and whole-machine robotics have no node, so they sit elsewhere in the atlas. Bubble area is market cap, ring weight is valuation premium, a dashed ring means currently loss-making.
Signal chain Power chain Actuation chain Bubble area ∝ market cap

Why the narrow test matters. It excludes much of what a looser “AI beneficiary” screen sweeps in, and produces a list you can argue with company by company. Coordinates are maintained by hand — process node and track tier have no field in the quant system — while size, ring weight and line style run off the daily data. Full stage-by-stage exposure table in the semiconductor coupling-point atlas.

//LINE A · VENTURE

We invest where the coupling points form.

A coupling point is a company that uses a signal to control power — where the compute stack meets the physical world.

FOCUS 01

AI silicon & interconnect

Accelerators, optics, memory, packaging.

FOCUS 02

AI software infrastructure

Where an architecture choice becomes a unit cost.

FOCUS 03

Robotics & embodied AI

Actuation, transmission, motor control, perception.

FOCUS 04

Frontier AI labs

Research running ahead of the product cycle.

WHAT WE BRING TO A CAP TABLE

  • An operator's read. Our founder built and shipped these systems.
  • Published coverage. Founders inherit an analyst who has done the reading.
  • Institutional reach. Standing bank and VC/PE relationships.
Seed → growth Silicon Valley led

//LINE B · RESEARCH

Research that banks read.

One technology at a time, worked down to the physics, the bill of materials, and the companies that capture the margin.

STREAM 01

AI & hardware deep dives

Model architecture read against the silicon it runs on — and what that does to unit inference cost.

STREAM 02

Industry & supply chain

Who supplies whom, where the single points of failure sit, which listed names carry the exposure.

STREAM 03

The Bit‑Watt wiki & atlas

A standing framework, not a one-off report — refreshed on the same daily cadence as the engine.

15

published deep dives

66

companies named in the atlas

15

sector verticals with valuation ceilings

219

stocks on the daily attention list

//LINE B · DEMO

What a deep dive actually looks like.

Research #15, on DeepSeek V4. Switch the GPU and watch where the workload lands.

Roofline: where MHA and MLA sit on the same hardware
Research #15 · vendor spec sheets · log–log
Diagonal is the bandwidth limit, plateau the peak-FLOPS limit — the two dots hold still while the roof moves.
Hardware

MHA achievable TFLOPS, batch 1

MLA achievable TFLOPS, batch 1

ridge point, FLOPs per byte

Only the roof moves. Operational intensity is an architecture constant — 1.2 FLOPs/byte for MHA, 55 for MLA — which on H100 is enough to lift batch-1 decoding off the bandwidth slope. Full derivation in Research #15.

//LINE C · QUANT DATA

A daily engine, not a monthly opinion.

Every number computed from raw market data inside our own pipeline. The universe is rebuilt after each close — running since December 2024.

5,203

A-share stocks rebuilt daily

170

dimensions per stock

298

dimensions project-wide

219

stocks on the attention list

36

US technology stocks tracked

THREE FACTOR FRAMEWORKS

  • Energy dynamics. Half-life decay over accumulated momentum.
  • Spring kinetics. Energy stored in consolidation, release calibrated per stock.
  • W-phase detection. Five-phase waveform classification with per-phase statistics.

THE DAILY PIPELINE

15:30Close triggers ingestion — OHLCV, margin, institutional flow.
16:00Per-stock profile rebuilt, factors scored.
17:00Sector matrices, macro heat map, crash-watch gates.
17:30Delivery — API, dashboards, alert push.

//LINE C · DEMO 1

170 dimensions in. One factor call out.

Dimensions are raw material; factors are how they become a decision. Two factor series below, drawn on three names — the same construction runs on all 5,203.

Factor output — energy and price position, last 30 sessions
A-share daily pipeline · as-of —
E(t), accumulated momentum after half-life decay — zero is the boundary. r60, price over its 60-day mean — 1.00 is the second. Together they resolve to one of four phases.

Two factors out of a 170-dimension profile. Each card shows the same pair computed the same way; only the phase they resolve to differs. The chip is the position rule the engine currently applies. Backtested returns are deliberately not shown — a backtest is not a deliverable, and it is not what a subscriber receives.

//LINE C · DEMO 2

Where the money actually goes.

Daily turnover across the 62 technology sectors we track under the Bit‑Watt framework, ranked every session. Two and a half years of Chinese technology capital rotating — press play.

Bit‑Watt sector rotation — top 30 by daily turnover
daily turnover, CNY 100M · —
Bar length is turnover on an absolute scale. Colour is fixed per sector; the top ten are drawn at full strength, ranks 11–30 dimmed.
Bit‑Watt turnover · of all A-share turnover From of 5,203 listed stocks sectors tracked · top 30 shown

The taxonomy is the product. These are not exchange classifications — the 62 sectors are built around supply-chain position under the Bit‑Watt framework, which is why PCB for AI server backplanes is separate from generic electronics. The board shows the thirty largest by turnover each session, drawn from a candidate pool of the 40 largest over the full period — wider than the display, so a sector that spikes enters the board that day rather than being locked out by its full-period rank. The turnover and stock counts above cover all 62.

//LINE D · ENTERPRISE · LIVE: A-SHARE

The layer institutions rarely build.

Market data is easy to buy; the layer above it is not. Deep technology research wired into a quant engine, so the research shapes the dimensions and the dimensions test the research. That join is what gets licensed, in five parts.

What follows is the A-share instance — the part that is live today, running daily since December 2024. The same five parts for US markets are in build; until they ship, US technology is covered on the research side.

Shanghai · Shenzhen · STAR Market 5,203 stocks, rebuilt every close US enterprise in build

01 · SENTIMENT

3 + 1

news feeds, plus the SOX index

Technology intelligence

Three Chinese market news feeds plus an overnight read-across from the Philadelphia Semiconductor Index, parsed daily by an LLM — yielding sector-level change, new companies for the coverage list, and alerts when a sector and its stocks move together.

02 · MACRO

95

market-level fields

Market conditions

ETF momentum, margin balances, block-trade flow, index trend and risk — the state the whole market is in, not one stock at a time.

03 · CRASH WATCH

19

gate dimensions

Drawdown early warning

Layered gates over sector price stress, computed daily. A watch state, an alarm state, and a severity score for episodes already under way.

04 · SECTOR HEAT

62

Bit‑Watt sectors

Concentration & rotation

Daily turnover, share and rank per sector — where capital actually is, on a taxonomy built around supply-chain position rather than exchange labels.

05 · PER-STOCK

170

dimensions per stock

The ticker layer

Price position, margin leverage, block-trade aggregation, fundamentals and chip structure, rebuilt for all 5,203 names after every close.

MODEL 01 · MOST FLEXIBLE

Custom API

Dedicated endpoints shaped to your consumption pattern.

REST · WebSocket · SFTP · S3 · backfill · uptime SLA

MODEL 02

Technology licensing

The five parts running inside your own infrastructure.

On-premise or private cloud · source-available factors

MODEL 03

White-label

Our dashboards and reports carrying your brand and domain.

Full UI rebrand · generated reports · sub-licensing

Who this is for. A smaller institution has neither a technology research desk nor a quant engineering team, and building either is a multi-year hire. A larger one usually has both — but separately, so the research never reaches the data dimensions and the data never carries the research’s taxonomy. The five parts above are already joined — today over A-shares, with the US build following the same five-part shape. US technology already runs as a 36-company attention universe on the research side, which is where that build starts from.

//LINE D · DEMO

Same engine. Same numbers. Your brand.

Pick a brand. Every figure below stays identical — the engine is what is licensed, the surface is what is yours.

White-label surface
illustrative brands, not clients
Only the branding is a setting. The stocks, the fields and the delivery time are not.
I
InAI Capital
Daily technology signal desk
quant.inaicapital.com
5,203STOCKS REBUILT DAILY
219ATTENTION LIST
170DIMENSIONS PER STOCK
17:30DELIVERY COMPLETE
TOP SECTORS BY TURNOVER CNY 100M VS 5 SESSIONS AGO SHARE OF BIT-WATT

Sector turnover is the sum of its member stocks' daily turnover — no price aggregation is involved, and the change column compares it with the session five trading days earlier. Same snapshot as the two demos above, . Nothing on this panel is mocked up.

Rebranding changes what people see, not what they get. Your logo, colour, product name and domain are settings. Which stocks are covered, which fields are computed and what time the file lands are the same engine we run for ourselves. A sub-licence lets you sell it on to your own clients, under your own contract.

//PARTNERSHIP MODELS

How we work together.

Whichever line you come through, the first step is a small one and the second is ours. If none of these rows describes you, write anyway — we would rather hear your question than have you pick a category.

Counterparty map
your step · then ours
IF YOU ARE THE LINE WHENEVER YOU ARE READY THEN WE
A startup or research lab A · Venture Send whatever you would normally send — a deck, a demo, a paper. No format required. Read it properly before we meet, and come back to you as soon as we can — including if it sits outside our focus.
An investment bank or research desk B · Research Tell us the technology or supply chain your desk needs covered. Send back what we have already published on it, and a scope if a commissioned deep dive would serve you better.
An asset manager or family office C · Quant Tell us how you would use the data — a universe, a horizon, a decision it feeds. Open trial access and walk the dimensions with you, so you can judge it on your own names.
A brokerage, data vendor or platform D · Enterprise Share the universe, the fields, and whose brand it ships under. Come back with a licensing shape and a pilot scope you can put in front of your own committee.
Media, a conference or an academic group A–D · all four Send the date, the audience, and the question you would like answered. Shape the material around it — hardware economics, the framework, the data, or the venture view, whichever actually fits.

//NEXT STEP

Tell us which line you need.

One address reaches all four.

CONTACT

[email protected]

Partnership, licensing, research and investment enquiries.

Entity

InAI Capital Advisor LLC
Silicon Valley, California

USEFUL TO KNOW — NONE OF IT REQUIRED

  • Which line you are writing about — or simply what the situation is.
  • The technology, universe or company on your mind.
  • Whether you are after research, data, a licence, or capital.
  • Your timeline, if there is one.

A few lines are enough to start. Every email is read and answered by us rather than a form — in English or Chinese, across US and China market hours.

Scope. InAI Capital Advisor LLC publishes research and quantitative data for information and education. It does not solicit funds, manage client capital, or provide investment advice, and nothing here is an offer to buy or sell any security. Direct investment activity sits with InAI Capital and is separate from the research and data business. Quantitative figures are a snapshot and change daily.