Home
Connexion
S'inscrire
Contenu d'apprentissage
Loading...
The $15,000 AI Bill. Your $20 Subscription is a DELUSION - Video học tiếng Anh
Pratique d'écoute
Pratique d'écoute
/
Video
/
The Infographics Show
/
The $15,000 AI Bill. Your $20 Subscription is a DELUSION
The $15,000 AI Bill. Your $20 Subscription is a DELUSION
Choisir le mode d'apprentissage:
Voir les sous-titres
Choisir le mot
Réécrire le mot
Highlight:
3000 Oxford Words
4000 IELTS Words
5000 Oxford Words
3000 Common Words
1000 TOEIC Words
5000 TOEFL Words
Sous-titres (486)
0:00
You think your $20 AI subscription is
0:02
the deal of the century. In reality,
0:05
it's a trap. A power user on tools like
0:07
Claude Code actually costs $15,000 a
0:10
year to run. But you're only paying a
0:12
fraction of that because venture
0:14
capitalists are footing the bill. You're
0:17
living inside the AI Uber moment, a
0:19
temporary illusion built to get you
0:21
hooked before the price tags change. But
0:24
the money is running out. When this
0:26
trillion dollar house of cards
0:27
collapses, the tools you rely on every
0:29
day will either vanish or cost you 10
0:32
times more. The economics of AI are
0:35
broken. Chapter 1, the $20 illusion. It
0:38
all starts with your wallet. A serious
0:40
Claude Code user runs through roughly 10
0:42
billion tokens a year. Tokens are
0:44
basically the thought units of AI. Every
0:47
word it reads, every word it writes,
0:49
every decision it makes relies on a
0:51
token. If you paid for that usage
0:53
through a standard API, those 10 billion
0:55
tokens would cost you around $15,000 a
0:58
year. That is the real unsubsidized
1:01
price. No discounts, no incentives, just
1:03
the raw compute costs. Now, that same
1:06
user on a flat rate max subscription
1:07
pays around $1,200 for an entire year
1:10
for the same workload from 15,000 down
1:13
to 1,200.
1:15
A 92% hidden subsidy. Imagine walking
1:19
into a dealership, picking out a car
1:20
priced at $15,000, and being told that
1:23
you only owe $1,200 because someone
1:26
somewhere else covered the rest. It
1:28
doesn't make sense, and that's what
1:30
makes this model so strange. But the
1:32
answer lies in OpenAI's own financial
1:34
projections leaked to the information.
1:37
The company is on track to lose $14
1:38
billion in 2026. Not revenue, losses. A
1:43
$22 monthly subscription covers about
1:45
1.7% of what an active power user
1:48
actually costs to serve. You are not a
1:50
customer. You are bait. Every prompt
1:53
typed, every line of code generated,
1:55
every late night chat session is being
1:57
paid for by investors and they are
1:59
betting that nobody will be able to live
2:01
without this product when the real bill
2:03
finally lands. Whole industries are
2:06
being signed up at a loss. Law firms
2:08
running document review at 5 cents on
2:10
the dollar. Marketing agencies are
2:12
turnurning out campaigns at prices that
2:14
would have been impossible 18 months
2:15
ago. Hospitals triing diagnostic tools
2:18
at sticker prices that no model provider
2:20
could actually sustain at scale. Every
2:23
single deal is being propped up by
2:24
patients capital that expects 10 times
2:27
returns. If companies are losing money
2:30
on every user they sign up, why are they
2:32
racing to sign up more? Because we have
2:35
seen this exact playbook before and we
2:37
know how it ends. Chapter 2. The ghost
2:40
of Uber. Back in 2014, a black SUV would
2:43
pull up outside your apartment in 3
2:45
minutes. The driver was polite, the car
2:48
spotless. The trip to the airport cost
2:50
you 11 bucks. You would wonder how any
2:53
of it added up. It didn't. And that was
2:55
the point. It was never meant to. For
2:57
the better part of a decade, an entire
3:00
generation lived inside what economists
3:01
later called the Millennial Lifestyle
3:04
Subsidy. Venture capitalists poured
3:06
money into ride sharing, food delivery,
3:08
co-working spaces, and meal kits on
3:11
purpose. They set the prices below cost
3:13
to crush legacy competitors and build a
3:16
habit. The plan was to take over first
3:18
and then raise prices until it made a
3:20
profit. Uber's take rate, the slice of
3:23
every fair a company keeps, tells the
3:25
story. In 2022, Uber kept around 32
3:28
cents of every dollar a rider paid. By
3:30
2024, that figure had climbed to roughly
3:32
42 cents. Drivers got a smaller share.
3:36
Riders paid more. The company eventually
3:38
posted a profit. Now it's happening in
3:40
the AI sector. It's the same investors,
3:42
the same playbook, and the same pricing
3:44
memo. Industry analysts expect consumer
3:47
subscription tiers to roughly double in
3:49
price over the next 2 years. Anthropic
3:51
has rolled out new rate limits that
3:53
gently push power users toward higher
3:55
priced plans. Google is testing premium
3:58
only Gemini features that used to be
4:00
free. A 100% price hike isn't a rumor.
4:03
It's already penciled in on the
4:04
calendar. Enterprise contracts are
4:06
following the same curve. Custom deals
4:08
signed in 2024 are being quoted much
4:10
higher in 2026 renewals. It's the same
4:13
product. It's just costing multiple
4:16
times the price. Users need to take it
4:18
or leave it. Ride sharing only had to do
4:20
one thing. Move a car from point A to
4:22
point B. The cost of doing that doesn't
4:24
explode as usage rises. If anything, it
4:27
gets more efficient. More drivers, more
4:29
density, better routing. AI works
4:31
differently. The underlying math of
4:33
thinking doesn't get cheaper in the same
4:35
way. It gets complicated fast. AI
4:38
executives continue to say that compute
4:40
is getting cheaper every year. The unit
4:42
economics will work out over time. It's
4:44
not exactly a lie. It's more like a
4:46
halftruth. The price of running a query
4:48
through a model has dropped
4:50
year-over-year. Chips are more
4:52
efficient. Models are leaner. Each
4:54
individual word an AI generates is
4:55
genuinely cheaper to produce than 18
4:57
months ago. And that's the part they
4:59
want people to hear. Here's the part
5:01
they don't. Chapter 3, the Claude code
5:04
math. Modern agentic workflows, the kind
5:07
that power Claude code and chat GPT's
5:09
deep research tools, burn through
5:11
anything from 5 to 30 times more tokens
5:13
than simple chat sessions of 2 years
5:15
ago. When you ask a code assistant to
5:17
fix this bug, it doesn't write 50 words
5:20
of response. It quietly spawns subtasks.
5:22
Then it rereads your files. It checks
5:24
its own work. It writes draft after
5:27
draft. Throws most of them away. and
5:29
then quietly runs tests in the
5:30
background. A single user request can
5:32
chew through hundreds of thousands of
5:34
tokens before any answer shows up. A
5:37
model might be slightly cheaper per word
5:39
than before, but it's also producing far
5:41
more words per request. The total bill
5:44
is shooting upward. It's known as the
5:46
token tax. It bankrupts scrappy AI
5:48
startups burning through their seed
5:50
rounds. It's threatening to wipe out one
5:52
of the most profitable business models
5:54
in the history of the internet. Chapter
5:56
4, the search penalty. For 25 years,
5:59
Google's printed money, and it's been
6:01
brutally simple. A user types in a
6:03
query, Google returns 10 blue links
6:05
pulled from the open web. The total cost
6:07
to Google, servers, electricity,
6:09
indexing is a fraction of a cent per
6:11
search. And yet, the ads next to those
6:14
results generate much more than that.
6:16
Margin is one of those great financial
6:18
miracles of modern times. Now, Google is
6:21
rebuilding that entire system on top of
6:23
generative AI. A single AI powered
6:26
search response, the kind that writes a
6:28
paragraph long answer instead of just
6:29
showing you some links, costs
6:31
significantly more to produce than a
6:32
traditional keyword search. Now multiply
6:35
that across billions of queries a day.
6:37
If Google fully replaces traditional
6:39
search with AI overviews, the most
6:42
reliable profit machine of the 21st
6:43
century vanishes. The margins that have
6:46
funded YouTube, Android, Whimo, and
6:48
Gmail begin to dry up. Wall Street
6:50
analysts have quietly mapped out the
6:52
worst case scenarios. And the numbers
6:54
are catastrophic. And it gets worse. The
6:57
advertising models become redundant,
6:59
too. When AI just gives you an answer,
7:01
nobody clicks on the links, so
7:02
advertisers will stop paying. Google is
7:05
staring at a future where it serves up
7:06
more queries than ever before, costs
7:09
more to run than ever before, and earns
7:11
less revenue per query than at any point
7:14
in its modern history. Tech giants are
7:16
willingly cannibalizing their most
7:18
profitable businesses on purpose.
7:20
They've decided the only thing more
7:22
dangerous than killing a cash cow is
7:24
letting a competitor kill it first.
7:26
Business school has a name for this, the
7:28
innovator's dilemma. When a new
7:30
technology threatens the core business,
7:32
incumbents face two choices. sit still
7:35
and defend the existing cash engine
7:36
while a competitor builds the future or
7:39
cannibalize it themselves on their own
7:41
terms, hoping that they can build
7:43
revenue on the next platform before the
7:45
old one erodess. That's the path
7:47
companies like Google, Microsoft, and
7:48
Meta are effectively betting on with AI.
7:51
They're betting that AI will eventually
7:53
replace the current money makers. Nobody
7:55
can prove that's true. Everybody is in
7:57
too deep to back out. If unit economics
8:00
are this bad, how are these same
8:02
companies posting record AI revenues on
8:04
Wall Street every single quarter?
8:06
Chapter 5, the roundtrip scam. That's
8:09
where things get clever. Microsoft
8:11
commits very publicly to investing $13
8:14
billion into OpenAI. The press release
8:17
is slick, the headlines dramatic, stock
8:20
prices rise. It makes investors happy.
8:23
But read the fine print and a different
8:25
story shows up. A big chunk of that
8:27
investment never actually hits OpenAI's
8:30
bank account. It arrives in the form of
8:31
Azure cloud credits. It's essentially a
8:34
gift card that can only be redeemed at
8:35
Microsoft's own data centers. OpenAI
8:38
records that sum on its balance sheet as
8:40
capital raised. Microsoft logs the cloud
8:42
usage as revenue. It's an investment and
8:45
a sale at the same time. Open AAI has
8:48
separately committed to spending up to
8:49
$250 billion on Azure services, locking
8:53
the loop in for years to come. Now layer
8:56
Nvidia on top of that. Nvidia announces
8:58
tens of billions in commitments to
9:00
OpenAI. OpenAI then turns around and
9:03
uses that capital to buy Nvidia GPUs.
9:06
Nvidia's quarterly revenue posts a
9:08
record and their stock price source. The
9:10
whole cycle takes a few months and
9:12
almost no real money has actually
9:14
changed hands. It has simply been given
9:16
a different name at each stop. Add
9:19
Oracle, Coreweave, and AMD to the list.
9:22
Each company invests and then sells
9:24
services to the next and records revenue
9:26
as the same dollar flows through the
9:28
cycle. The technical name for this is
9:30
round tripping. In Silicon Valley, it's
9:32
called strategic partnership. Chapter 6,
9:35
the hardware debt trap. In 2025, big
9:38
tech is projected to spend roughly 320
9:41
to$400 billion on AI infrastructure.
9:44
Updated forecasts for 2026 push that
9:46
figure toward 500 billion. data centers,
9:49
GPUs, cooling systems, power delivery,
9:52
entire grids are being reinforced to
9:54
handle it. Meanwhile, total global
9:56
consumer spending on AI services is only
9:58
about 12 billion. According to Menllo
10:01
Ventures State of Consumer AI report,
10:04
hundreds of billions are flowing out
10:06
while only 12 billion going in. The gap
10:08
is the size of an entire midsized
10:10
country's economy. It's being filled not
10:12
with revenue, but debt, corporate bonds,
10:15
structured credit, and private lending.
10:17
Meta alone raised $30 billion in bond
10:20
markets in late 2025. There was another
10:22
roughly $30 billion through a Morgan
10:24
Stanley arranged joint venture set up to
10:26
keep liabilities off of Meta's public
10:28
balance sheet. Microsoft has signed a
10:30
20-year power purchase agreement to
10:32
restart 3M Island. Google has partnered
10:35
with Next Era Energy to reopen nuclear
10:37
power plants. These promises don't go
10:39
away if AI revenue underperforms, but
10:42
the hardware itself doesn't last. A
10:44
high-end Nvidia GPU that powers most of
10:47
this boom has a short life of just 1 to
10:49
3 years before the next generation makes
10:51
them outdated. It loses most of its book
10:54
value the moment a new generation hits a
10:56
market, which now happens roughly every
10:59
18 months. A data center full of
11:01
three-year-old chips is in industry
11:03
terms dead weight. Compare that to the
11:05
original.com bust. When that bubble
11:08
popped in 2000, telecom companies left
11:10
behind millions of miles of fiber optic
11:12
cable buried in the ground. New
11:14
companies bought it for pennies on the
11:16
dollar and built YouTube, Netflix, and
11:18
Spotify on top of it. The crash was
11:21
brutal, but the wreckage was useful.
11:24
This AI bubble will leave behind
11:25
warehouses full of useless silicon,
11:27
locked up into 20-year power contracts
11:30
and concrete shells in the middle of
11:32
nowhere. No one will know what to do
11:34
with them. Utilities will pass higher
11:36
electricity rates on to the households
11:38
for decades, no matter whether the AI
11:40
revenues show up. A gap of hundreds of
11:43
billions of dollars cannot be papered
11:44
over for long. Companies running this
11:47
race already know it, so they're quietly
11:49
taking steps to slow the bleeding before
11:51
the public catches on. Most of the users
11:53
have already felt it. They just haven't
11:55
connected the dots. Chapter 7, the
11:58
stealth nerf. An AI model used to
12:00
oneshot your code. Now it forgets your
12:02
project halfway through. A chatbot used
12:04
to write five paragraphs a stretch. Now
12:06
it cuts off at three. An image generator
12:09
that used to render a flawless portrait
12:10
in 30 seconds now spits out something
12:12
with seven fingers and it asks for an
12:15
upgrade to the next tier. Nobody's
12:16
imagining these things. The product is
12:18
getting worse. When the numbers stop
12:20
working, the easiest lever a provider
12:22
can pull is to quietly water the service
12:25
down. The signs are easy to spot.
12:27
Message caps that used to refresh every
12:29
5 hours suddenly refresh every 8. The
12:31
default model in an app gets quietly
12:33
swapped from a flagship to a smaller,
12:35
cheaper version. Memory features get
12:37
rolled back. Advanced reasoning gets
12:40
locked behind a higher price tier. A god
12:42
model promised in launch keynotes is
12:44
quietly being swapped out for a cheaper,
12:47
less intelligent version. Reddit threads
12:49
about AI tools are full of users who
12:51
swear their assistant has gotten lazier.
12:53
Engineers are posting sideby-side
12:55
screenshots showing the same product
12:57
producing visibly worse output than 6
13:00
months earlier. Companies almost always
13:02
deny it. Sometimes they'll release
13:04
selected benchmarks, clean prompts,
13:06
controlled conditions, optimized
13:08
scenarios designed to demonstrate
13:10
performance at its best. It buys them
13:12
some time, but it doesn't fix the bigger
13:14
problem. A deeper issue has already
13:16
started taking out the first wave of an
13:18
entire AI ecosystem. Chapter 8, the 2026
13:21
mass extinction. Roughly 40% of AI
13:25
startups launched in 2024 have already
13:27
been shut down or aqua hired by bigger
13:30
players according to CB Insights data.
13:32
That is the polite term for a fire sale
13:35
where a struggling company is sold for
13:36
cents on a dollar to a rival. The buyer
13:39
isn't really buying a business. They're
13:41
getting the engineers shutting down the
13:42
product and absorbing whatever talent
13:44
they can absorb. These weren't hobby
13:46
projects in someone's garage. These were
13:48
companies that closed series A rounds
13:50
with serious investors. They had
13:52
revenue. They had paying customers. They
13:54
had glowing tech crunch profiles. Then
13:56
within 18 months, the lights went off.
13:58
The reason is almost always the same.
14:00
Their cost of goods sold, the money they
14:03
pay to model providers like OpenAI,
14:04
Anthropic, and Google is so high it
14:07
wipes out any margin they could hope to
14:09
charge. A startup that wrapped a
14:11
polished interface around GPT4 might
14:13
charge 50 bucks a month, but the API
14:15
usage that the same customer generates
14:17
can cost the startup $80. Every active
14:20
user is negative revenue. The more
14:22
successful marketing, the faster a
14:24
company bleeds out. When a foundation
14:26
model provider releases a new feature,
14:27
it often kills 10 startups overnight.
14:30
Chat GPT launches native voice mode. Say
14:32
goodbye to half a dozen voice agent
14:34
startups that closed series A rounds
14:36
last quarter. Claude releases native PDF
14:38
reading. A whole crop of document tools
14:41
became useless in a single product
14:43
update. An ecosystem of independent AI
14:45
companies is falling apart under the
14:47
weight of compute costs that nobody can
14:49
profitably absorb. When startups die,
14:51
cloud providers lose roundtrip revenue
14:54
that made foundation model investments
14:56
look like good business in the first
14:57
place. And that's when a final phase
14:59
begins. Chapter nine, the great AI rug
15:02
pull. Venture capital firms are no
15:05
longer willing to cover losses in the
15:06
hope of future glory. They want to see a
15:09
path to profit in writing with quarterly
15:12
milestones. and they want to see it.
15:13
Now, for foundation model companies,
15:16
that means one of two things. The first
15:18
is a brutal sudden repricing. A $20
15:20
consumer plan becomes a $100 plan, or it
15:23
quietly disappears and is replaced by a
15:25
protier that costs 10 times more for the
15:28
same features. A Claude Code user who
15:30
paid $1,200 a year suddenly faces an
15:32
invoice closer to $15,000 that an API
15:36
actually costs. A freelance designer who
15:39
relies on a $10 image generation
15:41
subscription gets an email explaining
15:43
that their plan is being moved over to a
15:45
new structure. Small businesses that
15:47
built workflows on cheap AI face a
15:49
choice. Pay 10 times more or go back to
15:51
doing it the old way. The second option
15:54
is worse. The services simply get shut
15:56
down. We've already seen the first
15:58
signs. Smaller AI companies have folded
16:00
with 30 days notice, leaving customers
16:03
scrambling to move years of work to
16:05
whatever competitor is still standing.
16:07
Specialized models for legal research,
16:09
medical imaging, and customer support
16:11
have been pulled because their economics
16:13
never worked. An era of cheap AI ends
16:15
with a thousand small invoices, a
16:17
thousand small shutdown notices. A
16:19
deeper truth is uglier than a price
16:21
hike. AI in 2026 is on track to become a
16:24
luxury, not a basic product. The cheap
16:26
versions trained an entire generation to
16:29
need it. An expensive version is the
16:31
only one that balance sheets now allow
16:33
to exist. Big companies that can afford
16:35
a new pricing tier will lock in their
16:37
advantage. Freelancers, the small
16:39
businesses, and the people who powered
16:41
early adoption, the ones who created the
16:44
buzz, will be priced out first. An
16:46
economy built on the idea of cheap
16:48
intelligence is about to slam into the
16:49
reality of expensive intelligence.
16:52
Productivity assumptions made in 2024
16:54
will not survive in 2027. A promised AI
16:57
revolution will arrive, just not for
16:59
everyone, and not at the price they were
17:01
sold. History says crashes don't take a
17:04
year to play out. The dot bust took 2
17:07
years from peak to trough. The AI bubble
17:09
has more leverage, more concentration,
17:11
and more debt baked into its
17:13
foundations. When it tips, it can move
17:16
in months, maybe weeks. When the margins
17:18
shrink, when the first big enterprise
17:20
customer publicly walks away from a
17:22
renewal, that confidence can vanish
17:24
overnight. The tools millions rely on
17:26
every day were never as cheap as anyone
17:28
thought. They were being held up by
17:30
investor money that is finally starting
17:33
to dry up. An AI age might still be
17:35
coming. A cheap AI age, one that fooled
17:38
an entire generation into rebuilding
17:40
their working lives on top of it, is
17:42
already over. A bill simply hasn't
17:44
arrived yet. And when it does, that
17:46
price will never feel real again. The
17:48
confidence that made the whole AI
17:50
industry feel inevitable is starting to
17:52
crack. What once looked like unstoppable
17:54
momentum is beginning to show the first
17:56
cracks of pressure beneath the surface.
17:58
Suddenly, the question shifts from how
18:00
big can this get to who is going to take
18:02
the hit when it doesn't. Find out in
18:05
what happens to the economy if the $2
18:07
trillion AI bubble bursts.
