Пузырь AI в 262728

Как строить продукт, если завтра тебя убьёт Claude / Codex

01
1977 → 2028
Platforms

Каждые 15 лет — новая платформа

Apple II · 1977

Personal computer

Bubble peak — 1983

Netscape · 1994

Internet

Bubble peak — 2000

iPhone · 2007

Smartphone

No bubble

ChatGPT · 2022

Generative AI

Bubble peak — 2028?

02
01 · Personal computer

Apple II — 1977.
Пик пузыря — 1983

The technology won. Most PC makers did not.

6 yearsfrom the first mass-market PC to the stock peak
Personal computers
06.1977Apple II — first mass-market PC
08.1981IBM PC
05.1983PC stocks peak
09.1983Osborne Computer goes bankrupt
~1987Apple back to its 1983 peak
03
02 · Internet

Netscape — 1994.
Пик пузыря — 2000

The network was built ahead of demand. Google and Amazon grew up on it.

−78%Nasdaq, March 2000 → October 2002
Nasdaq Composite
12.1994Netscape Navigator 1.0
08.1995Netscape IPO: +108% on day one
10.03.2000Nasdaq 5,048 — peak
09.10.2002Nasdaq 1,114 · −78%
04.2015Nasdaq back to the peak — 15 years later
04
03 · Smartphone

iPhone — 2007.
Пузыря не было

Built from profits, not debt. The money went to the two platform owners.

Applehardware margin + 30% of every App Store sale
Googlegave Android away free → ~72% of the world’s phones search with Google
Google → Apple~$20B a year (2022) to stay Safari’s default search
0bubbles: smartphones paid off from day one
Mobile
09.01.2007Jobs unveils the iPhone
07.2008App Store
2008Crisis — mortgages, not tech
08.2011Apple — world’s most valuable company
05
04 · Generative AI

ChatGPT — 2022.
Мы здесь

The fastest platform in history — and the most expensive build-out.

4 yearssince ChatGPT. The internet took 5 years to peak
Generative AI
30.11.2022ChatGPT
01.2023100M users in 2 months
2026900M weekly, 5% pay
12.11.2026Nvidia −30% from peak, Huang steps down
19.11.2026Anthropic IPO
06
Pattern
Rhythm

От продукта до пика — 5–6 лет

198019902000201020202030todayPC6 yrsInternet5 yrsSmartphoneno bubbleAI5–6 yrs?mass productbubble peak
2022+5–6=2028
07
Adoption
Adoption

ИИ приходит вдвое быстрее

Share of US adults using the technology

Generative AIafter 2 years
39.5%
Internetafter 2 years
20%
Personal computerafter 3 years
20%
08
Revenue run-rate
Speed of money

До $100 млн выручки — за 8 месяцев, а не за 10 лет

061218243036months from launch to $100M revenueLovableapps from a prompt$500M · 19 mo8HiggsfieldAI video$1B · 18 mo8ManusAI agent9 moReplit Agentapps built by an agent9 moGensparkoffice agent9 moCursorcode editor$4B · ~29 mo12ElevenLabsvoice AI20 moSlackclassic SaaS record30 moZoomvideo calls→ 54 moAverage SaaStop-70, from founding→ 120 mo
15×faster than the average top SaaS company
4×faster than Slack, the classic SaaS record
11.5 momedian time to $1M for Stripe’s top-100 AI companies
09
Valuation vs revenue

Оценки идут за выручкой

$300M$1B$10B$100B$1TOpenAIfrontier models, ChatGPT25×Anthropicfrontier models, Claude15×CursorAI code editor15×ElevenLabsvoice AI44×HarveyAI for law firms39×Lovableapps from a prompt27×HiggsfieldAI video8×
annual revenue (run-rate)last valuation× valuation ÷ revenue
Anthropic multiple
20× → 15×

May 2026 round → IPO price range. Multiples shrink as revenue arrives

Second tier

Products on top of models are priced at 8–44× revenue. Investors pay for revenue, not promises

10
The answer up front

Короткий ответ: пузырь есть, но не везде

Bubble — here
Дата-центрыBig-tech capex $147B → $730B a year, 2023 → 2026
Долг$1.5T of AI borrowing in three years
Цена входаStartups at 8–44× revenue
vs
Real — here
ВыручкаAnthropic run-rate
$0.4B$81B
2024 → October 2026
“Every bubble is different! But it can still be a bubble”Benedict Evans, 2025
11
01

Капитал

Кто платит за дата-центры

12
Capital · AI share of global VC

ИИ забрал венчур
и собрал его в мегараундах

18%202335%202450%202580%Q1 26>70%Q2 2664%Q3 26
43%
of global VC in H1 2026 went to OpenAI and Anthropic
Rounds under $100M: 44% → 12.5% of US deal value
13
Capital

Мои клубы: ИИ — треть
и рост остановился

1.3%20183.8%201911.2%20206.8%20218.8%202216.8%202321.4%202438.3%202537.8%2026
1 of 76
pitches were about AI in 2018
339 of 897
of the last pitches I reviewed were about AI
14
Capital

Расходы на дата-центры ×5 —
свободные деньги кончаются

67201895202015120221472023228202437620257302026 plan229.761.5
Capex, $B Free cash flow, $B
−73%
free cash flow of the big four since 2024
15
Capital

Разницу добирает долг
и деньги ходят по кругу

55%of AI money comes from AI
Nvidia
chips
OpenAI
models
Oracle · CoreWeave
data centres
$30B into the round
multi-year compute leases
buying Nvidia GPUs
$1.5T
AI-related borrowing over three years — Goldman Sachs
≈15%
CoreWeave bond yield
16
Capital

This is not 2000: capacity lags demand

Fibre · 2000
Ahead of demand
capacitydemandidle
Fixed costBuilt “for growth”Most fibre lay dark
AI compute · 2026
Behind demand
demandcapacityshortage
Cost per queryPaid by usageNot enough capacity
17
02

Это реально?

Есть ли выручка и польза

18
Is it real? · Anthropic revenue

Выручка настоящая —
у 2000-го её не было

$0.04B2023$0.4B2024$4.6B2025$33.2B9M2026$81Brun-rateOct 2026
$33.2B
Anthropic revenue in the first 9 months of 2026
19
Is it real?

ИИ делает
всё более длинные задачи

3.6 hGPT-55.3 hClaude Opus 4.514.5 hClaude Opus 4.6*16 h+Claude Mythos*
* reported · above 16 hours the benchmark runs out of tasks
≈4 mo
for the length of tasks an agent can do alone to double
20
Is it real?

Польза в задаче есть,
в экономике — пока нет

METR · experienced developers
+20%Felt faster−19%Measured
Controlled experiments
Support agents+14% Consultants, speed+25% Developers, Copilot+26%
Whole economy
Wages & hours, Denmark<1% US labour productivity+1.3% TFP gain over 10 years<0.53%
21
Is it real?

Шириной в милю, глубиной в дюйм

Mile wide · ChatGPT
900M

weekly users

5%

pay

80%

ask <3 questions a day

Inch deep · people who work in the terminal
0%25%50%75%100%2025202620272028forecastClaude Code 39%Codex 16%~70%?
Claude CodeCodexshare of developers using it at work
≈19M

developers on Claude Code (est.)

5M+

weekly Codex users, Jun 2026

$2.5B+

Claude Code run-rate, Feb 2026

OpenAI + Anthropic revenue vs. $1.5T of revenue needed per year of build-out
<10%
22
03

Рынок

Сколько ИИ уже в вашем портфеле

23
Market

Купили индекс — уже наполовину в ИИ

S&P 500
≈45%

AI-linked share of market cap

Magnificent 732.6%
Semiconductors5% → 15%
Goldman, 2022: 25%JPMorgan, Oct 2026: 49%
Nasdaq-100
≈60%

Mega-cap tech + chips

Magnificent 737.9%
Semiconductors23%+
Mag 7, July 2023~55%
Russell 2000
≈24%

AI-linked small caps

Share of H1 2026 gain52%
IT sector weight11.9%
75%

of S&P 500 returns since ChatGPT came from AI stocks

Shiller CAPE
avg 17now 40.21999: 44.2
24
Market

Anthropic IPO, Nov 19 —
an exam for the whole market

May 2026 round$965B IPO, basic shares$1.0T $70–78 a shareIPO mid, diluted$1.2T IPO top, diluted$1.4T If priced above range$1.5T September hype$2.0T not the terms
30%+
operating margin needed to justify the valuation (Damodaran)
20× → 15×
valuation ÷ revenue: the froth is coming off
25
Market

Политика стала риском

After the 4 November elections, data-centre construction became an issue for governors and Congress.

US planned data-centre capacity
40%

in states where new governors promise a pause

OhioGeorgiaMichiganNew MexicoTexasNew York*Oregon*

* extend pauses already in force

Public mood
64%

of Americans say AI is moving too fast

−12.3%

chip index (SOX) on 5 November, the day after the election

So what

Now politics — not only money — can stop the build-out.

26
Market · what I watch

Четыре датчика, которые покажут разворот

01 · Credit spreads
calmstress
≈15%

CoreWeave bond yield

02 · Capex plateau
growthplateau
+34% → −3%

capex growth, 2027 → 2029 consensus

03 · Politics
calmstop
40%

of planned capacity in states promising a pause

04 · Energy
enoughshort
220 GW

turbines on order vs 60–70 GW built a year

Fund managers expecting the AI bubble to burst in 2027 (BofA survey)
45%
27
Where profit goes

Модели дешевеют — прибыль уходит выше

Price of the same intelligence
2023202420252026100÷5–10 a year

schematic: cost of a fixed level of quality, indexed

value
Apps & servicesvalue moves here
Modelscommodity · same quality ÷5–10 a year
Cloud & data centrescapex
Chipsscarce today
Lesson of mobile networks

$1T revenue a year, $200B capex — and operator stocks went nowhere. The money went to those above them.

69%

of company spend on applied AI still goes to the labs — for now

28
2022 → 2030
Waves

Три волны ИИ-продуктов — и четвёртая впереди

2022
The race starts

ChatGPT and GPT-4. Big labs build the models

OpenAIAnthropicGoogle
2022 – 2025
Wrappers

A product on top of a model. Winners broke through and owned a channel

CursorPerplexityLovableMidjourneyElevenLabsGammaHiggsfield
2025 – 2027
Verticals

AI for accounting, logistics, sales, marketing. Already a red ocean

BasisRilletAugmentClay11x
2027 +
What will we still use?

Products that survive once models can do the job themselves

?
29
2022 – 2025 · Wrappers

Победили те, кто забрал свой канал

Cursor
code editor
channeldevelopers’ IDE
~$4B ARR · SpaceX buys for $60B
Lovable
apps from a prompt
channelevery app is an ad: 900M visits a month
$500M ARR · $13.3B
Midjourney
images and video
channelDiscord: ~20M members
no VC · ~$500M revenue (est.)
Perplexity
answers with sources
channelsearch habit + telecom bundles
>$750M ARR (reported)
ElevenLabs
voice AI
channelcreators and developers via API
>$500M ARR · $22B
Gamma
presentations
channelevery deck shown is an ad
$100M ARR · profitable
Harvey
AI for lawyers
channelsales to large law firms
>$400M ARR · $15.5B
Higgsfield
AI video
channelcreators and agencies
$1B run-rate · $5.4B → offers at ~$20B
30
Red ocean
2025 – 2027 · Verticals

В вертикалях океан уже красный

~50%of Y Combinator’s Summer 2025 batch build AI agents
Accounting
Basis — $1.15B, Feb 2026
Rillet — $1B, third round in a year
Numeric, Campfire, Accrual, Modus
Elad Gil: the niche is still open
Logistics
Augment — $85M Series A
HappyRobot, FleetWorks, Vooma, Pallet
$9B invested in 2025
Reuters: investors fear saturation
Sales
Clay — $7.1B
11x — customer-logo scandal
Artisan and dozens of AI SDRs
Deal count down 4 years in a row — money goes to a few
Marketing
Jasper — lost its market
Writer — $1.9B in 2024
HubSpot and Adobe built AI in
Platforms take the feature in-house
31
Entry price

Entry price decides everything

Miro
$17.5B$1.79B2021 roundsale, Sep 2026

revenue doubled, and it’s profitable

Character.AI
$1B$20M2023 valuationDisney bid, Oct 2026

bankruptcy, October 2026

Y Combinator 2025–26 · batch size · building agents (X25, S25) / AI-first (F25, W26)
166W2550%X2550%S2592%F2588%W26196X26235S26219F26
$500K

standard deal: $125K for 7% + $375K uncapped SAFE

$20M → $40M

typical Demo Day cap, W23 → W26; ~10 companies at $100M

“A great company and a great stock are different things”Patrick Boyle, Web Summit 2026
32
Portfolio

One deal is a lottery,
twenty is a strategy

15 of 100
reach Series A within two years (31 in 2018)
2–4 of 100
return the whole fund
Seed— yes, 20+ dealsSeries A— only with revenueSPVs into AI labs— no
33
Checklist

Three questions before the check in 2026

01
Domain expertise
years in the industryyears writing code with AI

The founder knows the industry from the inside — years in accounting, logistics, medicine. Not only code.

02
Distribution
CursorIDEMidjourneyDiscordLovableevery app = ad

Knows where the customer is and can own that channel. That is how Cursor, Midjourney and Lovable won.

03
The 2030 test
modelsyour moat20262030

Knows what happens to the product when models learn to do it themselves.

The real question: which products will we still use in five years?
34
Due diligence

Eight questions
to ask a founder before the check

01
What does inference cost per customer?
02
What if OpenAI ships this as a feature?
03
Can you switch to a model 10× cheaper?
04
How much revenue is contracts vs. pilots?
05
How many customers stay after 12 months?
06
What data do you have that competitors don’t?
07
What if there is no next round?
08
Why this valuation?
35
My view · November 2026

Где перегрето, а где ещё есть смысл

OverheatedI cut
GPU clouds on debt
SPVs into AI labs
Wrappers on someone else’s model
Series A with AI only in the name
Still makes senseI keep going
Vertical AI with its own data
AI inside a service with paying customers
Products that gain from cheaper models
Infrastructure for agents
temperaturecold → hot
36
Playbook

Как ангелу фиксировать прибыль

01
Sell 20–30% on secondaries

Don’t wait for the top. If a hot round lets you sell, sell some.

02
Pro-rata only in leaders

Where the unit economics already work — not where the story is loudest.

03
Rebalance AI back to target

In the public portfolio the index has already done the buying for you.

After the 2000 peak
−49%S&P 500
−78%Nasdaq

Keep a reserve for 2027.

37
My answer

Инвестировать или фиксировать? И то, и другое

Take profit
Public AI stockstrim, rebalance to target
Late rounds & SPVsdon’t buy at the labs’ prices
Keep investing
Early dealssmall checks, 20+ companies
Products above the modelsown channel, own data, gain from cheaper models
Five rules for 2027
  1. Check who actually pays
  2. Treat inference as COGS
  3. Look for what can’t be vibe-coded
  4. Syndicate into leaders, not VC funds
  5. Keep a reserve
So how do you build a product if Claude could kill you tomorrow?Here is ours →
38

Сделали для своих задач. Понравился не только нам

D.A.V.E. started as the assistant that runs our own companies. Then other founders asked for it.

Mac & Windowsruns on your Claude / ChatGPT subscriptiondave.space
39
What D.A.V.E. is

D.A.V.E. — Lovable для бизнеса

Lovable builds an app from a prompt. D.A.V.E. does the actual work: sites, dashboards, CRM, mail, reports — on your computer, in your tools, on a schedule.

Lovable
“Build me an app”→ an app
D.A.V.E.
“Run this for my business”→ the work, done — every day
Running in our companies right now
07:00Morning dashboarddata, payouts from 5 booking sites, deploy✓
08:00 · 19:00Industry newsUS + RU sources → brand dashboard✓
09:0010 improvements a dayproduct dashboard, shipped to production✓
dailySEO for 3 sitesedits, checks, deploys✓
on requestGuest reviewsAirbnb, Booking, HomeToGo — in the host’s voice✓
22:00Business idea a dayresearch, financial model, landing page✓
40
How it works

Не воюем с Claude — строим поверх

One desktop agent. The best engines underneath. One brain on top that remembers you.

ClaudeCodexKimi

switch engines in one chat — on your own subscription

↓
One brain for every engine
memoryskillsprojectsruleschats that hand work to each other
↓
filesbrowserappsservices · API · MCPscheduleTelegram
41
Dry numbers

Та же задача — до 4 раз меньше времени у экрана

Task: build a finance dashboard from HubSpot, office mail and 5 files — then act on it. Same model, Opus 5.5 High.

Human time at the screen
D.A.V.E.
6 min
Claude app
11 min
Hermes
11 min
Codex app
22 min

Less screen time is the whole point: the agent finishes the work, you approve.

What got doneD.A.V.E.Hermes
Finance dashboard✓✓
Supplier orders placed2—
Customer order cancelled✓—
Edits in books & CRM13—
Emails to customers74
Scheduled tasks21
D.A.V.E.
Hermes
42

Вопросы?

Roman LoginovTelegram @tarfoo
Telegram · @tarfoo
D.A.V.E. · dave.space
43