Home
Вход
Регистрация
The Leaked Audit That Exposes OpenAI's Real Financial Crisis - Video học tiếng Anh
Практика аудирования
Практика аудирования
/
Video
/
The Infographics Show
/
The Leaked Audit That Exposes OpenAI's Real Financial Crisis
The Leaked Audit That Exposes OpenAI's Real Financial Crisis
Выберите режим обучения:
Посмотреть субтитры
Выбрать слово
Переписать слово
Highlight:
3000 Oxford Words
4000 IELTS Words
5000 Oxford Words
3000 Common Words
1000 TOEIC Words
5000 TOEFL Words
Субтитры (245)
0:00
Sam Altman promised the world that AI was the ultimate cash machine.
0:04
But behind closed doors, OpenAI is the midst of a crisis. The company
0:08
is bleeding money. The biggest AI company in the world built a
0:11
product so revolutionary that their own ambitions are bankrupting them.
0:15
How did this happen… and what happens when the money finally runs out?
0:19
OpenAI made a promise to its investors that would make any sane person do a double take.
0:24
The company plans to keep posting massive annual losses all the way through 2028. Then, somehow,
0:30
it expects to suddenly become profitable, hitting positive cash flows sometime around 2029 or 2030.
0:36
Worryingly the promise keeps changing.
0:38
And always in the same direction.
0:41
Originally, the company told its investors the 2030 revenue would be close to $175
0:46
billion. But they would spend close to $200 billion getting there. By February 2026,
0:51
the revenue target had jumped to $280 billion. And the burn
0:55
estimate ballooned right along with it. It hit a massive $665 billion.
1:00
Every time the losses get bigger and the losses get bigger too.
1:04
Promises like that are just spreadsheets with good intentions. At least until someone gets
1:09
a look at the books. And in the summer of 2026,
1:12
someone actually did. A set of documents landed in the hands of a journalist named Ed Zitron.
1:18
Journalists are used to stumbling on big stories, whether through whistleblower
1:22
testimonies or leaked e-mails. This, however, felt like something more mundane: a set of OpenAI’s own
1:28
audited financial statements, documents the Financial Times confirmed were real.
1:32
How controversial could an audit get?
1:35
As it turned out, the documents revealed that in 2025, OpenAI’s revenue nearly quadrupled
1:41
to over $13 billion. That same year, the company posted a net loss of $38.5 billion
1:48
Somehow, OpenAI managed to lose more money in a 12 month period than it
1:52
had made in its entire history… combined. Sam Altman has spent three years telling
1:58
the world that AI is the most valuable business on the planet. The company he
2:02
runs, the one supposedly worth a trillion dollars, is bleeding out.
2:06
Whenever someone brings up OpenAI’s losses, the company has a go-to response.
2:10
They’ll talk about efficiency and improving ratios. They’ll say they’re
2:14
eating infrastructure costs upfront that will later pay for themselves many times over.
2:19
They’re not lying.
2:20
In 2024, OpenAI spent $2.25 for every $1 it made. In 2025, it dropped to
2:28
$1.60. It’s a better ratio, and it might look like the company is getting its act together.
2:34
It’s not.
2:34
The ratio story got put to the real test in early 2026. Documents revealed OpenAI’s revenue nearly
2:41
tripled year over year to $5.7 billion in the first 3 months of the year. Sounds great,
2:47
until you realize there was a $3.7 billion cash burn in the same quarter.
2:52
Over half of its net earnings.
2:54
Add that to the $9.3 billion operating loss for
2:58
those 3 months and the $12.4 billion non-cash charge tied
3:03
to the company’s corporate restructuring. The total comes to a $21.3 billion loss.
3:09
In one quarter.
3:10
CFO Sarah Friar has described this growth, in an actual company blog post,
3:15
as growth “never seen before at such scale.” Sure, the scale part checks out. But what
3:21
about the fact OpenAI’s own internal targets didn’t survive first contact with 2026 either?
3:27
The Wall Street Journal reported that OpenAI missed its internal projections for both
3:31
revenue and weekly active users in early 2026. It was the first time that had happened against
3:37
Friar’s forecasts. Weekly active ChatGPT users peaked around 920 million in February,
3:43
then dipped to an average of about 905 million for the quarter. Around that same time, Friar
3:48
was reportedly telling people internally that OpenAI might not be “IPO-ready”.
3:54
Something else changed.
3:55
Friar used to answer directly to Sam Altman. Reports claim she now
3:59
answers instead to Fidji Simo, OpenAI's CEO of AGI Deployment,
4:04
essentially getting a new boss placed between her and the CEO. Outside reporting has read
4:09
that as a sign of real friction inside the company over the burn-rate problem.
4:13
The company’s cost-per-dollar ratio really did improve between 2024 and 2025. But a company can
4:20
look more efficient on paper while its losses keep growing faster than that ratio can shrink.
4:25
OpenAI's own guidance admits that.
4:28
Their full-year cash burn for 2026 is projected near
4:31
$25 billion. It’s hovering near $57 billion for 2027, roughly doubling.
4:38
A better ratio wrapped around exploding dollar losses feels like the same problem.
4:43
The numbers just make it sound better.
4:45
The losses on their own aren’t even the strangest part of the
4:48
story. That’s buried in how they’re accounting for the actual computers.
4:52
When a company buys expensive equipment, like,
4:54
say, tens of thousands of Nvidia GPUs, it doesn’t count the whole cost as an
5:00
expense on day 1. It will spread that cost out over the equipment’s “useful life,”
5:04
or how many years it’ll realistically stay useful before it’s outdated junk.
5:09
Microsoft says its GPUs last for 6 years. Meta says 5 and a half.
5:14
Michael Burry, the person who predicted the 2008 housing crash thinks that’s nonsense.
5:19
His argument is that Nvidia drops a brand new chip every 12 to 18 months. So a GPU
5:24
you bought today is basically outdated in a year and a half, even if it physically
5:29
still works. His estimate for real useful life is 2.5 to 3 years, not four to six.
5:35
That matters.
5:36
A longer “useful life” number would mean smaller expense numbers showing up on the books at
5:40
the end of each year. That makes profits look bigger than they might actually be.
5:45
Burry says this trick, done across the entire industry, is hiding roughly $176
5:50
billion in real expenses between 2026 and 2028. That’s all the big players,
5:56
granted. But OpenAI is buying and leasing more chips than almost anyone.
6:00
Nvidia and CoreWeave say Burry’s wrong. To them, older chips are still worth real
6:05
money on resale and customer contracts run for 5 years anyway. Meanwhile,
6:10
Microsoft’s own CEO admitted they’re slowing down some data center construction. They’re
6:15
worried about overbuilding for hardware that’ll be outdated before the buildings even finish.
6:20
So who’s right?
6:21
Well, nobody knows, mostly because chips running this hot at this scale have never existed before.
6:27
There’s no 10 year track record to refer to. The accountants signing off on these numbers are,
6:32
best case, making an educated guess about a bill nobody’s actually seen yet.
6:37
Burry’s other comparison is Cisco during the dot-com crash.
6:40
Cisco’s stock lost about 80% of its value in 2 years because their spending
6:45
got way ahead of the revenue that was supposed to justify it. Just like then,
6:50
people are starting to wonder if the accounting departments at the heart of
6:53
the AI revolution might be painting a rosier picture than the actual machines can back up.
6:59
Every company that’s ever bought expensive equipment has had to take a gamble on how
7:03
long it’d last. Sometimes those guesses turn out to be more or less right. The
7:08
problem is the speed and scale of the AI enterprise. Hundreds of billions
7:12
of dollars are currently riding on a guess about technology that’s evolving
7:16
faster than any hardware category in modern history. Get that guess wrong,
7:21
even by a year or 2, and you're talking about insane losses across the entire industry.
7:26
But how is OpenAI funding all of this?
7:28
In September 2025, Nvidia announced it would invest up to $100 billion into OpenAI. Everyone
7:35
celebrated as Nvidia’s valuation jumped past $4.5 trillion in a matter of weeks.
7:41
But it was an illusion.
7:43
A circular loop of financial trickery.
7:45
Nvidia gives OpenAI money. OpenAI uses that money, plus other funding it has raised,
7:51
to sign huge cloud contracts with companies like Oracle.
7:54
Oracle then uses that money to go buy a ton of chips… from Nvidia.
7:59
So the cash starts at Nvidia,
8:00
gets labeled as “investment,” then moves through two other companies. Eventually,
8:05
it landed right back in Nvidia’s bank account, this time labeled “revenue.”
8:09
Every single transaction is real.
8:11
It’s all completely legal.
8:12
Nvidia gets to report record chip sales. OpenAI gets to announce that it’s
8:16
committed a huge chunk of cash to improving its own infrastructure.
8:20
And Oracle gets a guaranteed $500 billion of future business. 3 companies are helping each
8:25
other and pointing at the same pool of money, calling it proof that demand is exploding.
8:30
The model isn’t infallible.
8:31
By January 2026, cracks started to appear.
8:35
The Wall Street Journal reported the promised $100 billion had basically stalled out. Behind
8:39
closed doors, Nvidia’s CEO Jensen Huang reportedly told people the original deal was never actually
8:45
locked in. He allegedly had real concerns about OpenAi’s quote, “lack of financial
8:50
discipline.” To say nothing of his worries about competition from Google and Anthropic.
8:55
The number on the table suddenly dropped from $100 billion to something closer to $30 billion.
9:00
Meanwhile, OpenAi’s total compute commitments across Microsoft, Amazon,
9:04
Oracle, Nvidia, and AMD had stacked up to a reported $1.4 trillion. For
9:10
a company that made $13 billion total in 2025, that’s quite the pill to swallow.
9:16
On Wall Street the process is called vendor financing, or revenue-round tripping. It’s
9:21
the exact same move that inflated telecom stocks right before the dot-com bubble popped,
9:26
when equipment makers financed their own customers buying that same equipment. The
9:30
plan worked great until the demand that was supposed to manifest… simply didn’t.
9:35
Analysts flagged this exact circular concern the same week the original Nvidia deal was announced.
9:40
The alarming part is that every quarter, headlines will report that AI infrastructure
9:45
spending has hit record highs. Every quarter, these records get treated as
9:48
proof that demand for AI is exploding. But if a meaningful chunk of that record spending is
9:54
the same small circle of companies passing capital back and forth between each other,
9:58
then record spending and record demand all become relative terms from the start. Record spending
10:03
might be a fact, but record demand is a story that record spending is getting used to tout.
10:09
Nvidia is just one piece of this.
10:12
There’s a whole other loop running through Microsoft.
10:14
Microsoft has been backing OpenAI since 2019. The tech giant had put over $13
10:19
billion into the AI company before its giant funding rounds even started. In exchange,
10:24
Microsoft got to resell OpenAI’s models on Azure and its cloud computing platform.
10:29
For years, it paid OpenAI a cut of whatever it made doing that.
10:32
Roughly 70% of Microsoft’s entire reported AI revenue traces back to one single customer:
10:38
OpenAI. That number is what Microsoft executives use to
10:42
justify spending $190 billion a year on infrastructure alone.
10:46
Where does OpenAI get the cash to pay its Azure bill?
10:49
Mostly from the same investor pool, Microsoft included, who keep funding
10:54
OpenAI’s next round specifically so it can keep paying for the compute
10:58
that Microsoft then books as revenue. It’s that circular finance loop again.
11:02
Money leaves Microsoft; money lands in OpenAI’s hands;
11:05
a big chunk of money walks straight back to Microsoft as a cloud invoice; Microsoft
11:10
then shows shareholders this money as evidence its AI bet is paying off, and everybody wins.
11:16
In April 2026, Microsoft actually gave up part of this deal. They agreed to stop
11:22
taking a revenue cut on Azure payments and let OpenAI use Amazon and Google’s clouds,
11:26
too. It looked like OpenAI was finally carving out some long-overdue independence.
11:32
Except that only helps if there’s somewhere else to
11:35
get the truly absurd amount of cash OpenAI needs.
11:38
That’s where the problems really begin.
11:40
Every version of OpenAI’s long-term sales pitch eventually lands on one number:
11:44
$100 billion in annual revenue by 2030. That number used to be built on the assumption its
11:51
$20 per month ChatGPT subscriptions would represent the biggest chunk of its pie.
11:56
In April 2026, the company revealed that it is now counting on something entirely
12:00
different to reach those lofty projections: ads.
12:04
OpenAI made $2.5 billion in ad revenue in 2026. They project $11 billion in 2027,
12:11
rising each year before, somehow, hitting $100 billion in 2030.
12:16
It kind of feels like crowning yourself the winner of a Monopoly game you’ve just started.
12:20
For that math to check out, OpenAI needs to go from about 900 million weekly users today to 2.75
12:28
billion by 2030. Those are Facebook-at-it’s-peak numbers, a feat which took Facebook, Instagram,
12:34
and Whatsapp a combined 17 years to pull off… across all three ecosystems.
12:40
There are roughly 8.2 billion people on the entire planet, including children,
12:44
people with no internet access, and people who have never touched a computer. OpenAI
12:49
wants roughly a third of humanity to open their app every week in a
12:53
timeframe shorter than it takes most people to finish a college degree.
12:56
And the ad market it’s counting on might not even be big enough once it gets there.
13:01
Analysts at eMarketer estimate the entire U.S. chatbot ad market,
13:05
every competitor combined, at under $6 billion by 2030. OpenAI's own target for that same year,
13:12
just from its slice, is more than 16 times that estimate.
13:15
How, exactly, will they sell ads to a market that mostly doesn’t exist?
13:20
They won’t. Not unless something else changes drastically first.
13:24
And that something is the one thing no slick investor deck can conjure up.
13:28
Every time ChatGPT answers you it runs on a GPU that’s purpose built for punishing computation.
13:34
Nvidia’s current generation of H100 chips draw about 700 watts on their own. But these GPUs
13:40
never run by themselves. Slot it into a server with the CPUs, memory, and networking hardware
13:46
it actually needs to function, and the real draw per GPU climbs to roughly 1,300 watts.
13:51
Nearly half of that is overhead the chip itself never sees.
13:56
When you scale that number up to an actual training cluster the numbers
13:59
get pretty eye-watering. A facility running 100,000 H100-class GPUs will need somewhere
14:05
around 200 megawatts pulled from the grid to compensate for raw chip draw, networking,
14:11
cooling, and power-delivery. Over a year, that’s enough energy to power close to 165,000 homes.
14:18
That’s just today’s hardware.
14:20
Nvidia’s next architecture is expected to draw more than 3x what a single H100 does,
14:26
per GPU. So the power bill isn’t shrinking anytime soon.
14:30
What OpenAI is building is far more ambitious than a couple of isolated
14:34
clusters. They’re intent on constructing something called Stargate. This joint
14:38
venture with SoftBank and Oracle will spend up to $500 billion on data centers over 4 years.
14:44
That’s on top of a separate $250 billion Azure commitment running through 2032.
14:50
Total projected compute spend through 2030 will come to somewhere around around $600 billion.
14:56
OpenAI made $13 billion in 2025.
15:00
That gap is exactly why OpenAI closed a $122 billion funding round back in March of 2026,
15:07
valuing the company at $852 billion.
15:11
What’s buried in the fine print of that deal?
15:13
Well, for one, the fact that Amazon anchored it with a $50 billion commitment. But $35 billion
15:19
of that only shows up if OpenAI either goes public by the end of 2028 or achieves artificial general
15:25
intelligence. One depends on regulators and bankers cooperating on a specific timeline;
15:31
the other depends on OpenAI inventing something no lab on Earth has ever built.
15:35
In March 2026, a planned 2-gigawatt expansion of the Stargate site in Abilene,
15:40
Texas got canceled. That one project alone would have powered almost 1.5 million homes,
15:46
and it got scrapped from a buildout that’s supposedly racing towards $600 billion.
15:50
New data center power connections in places like Northern Virginia now take
15:55
4 to 7 years to get approved. That means that some of the power OpenAI's counting
15:59
on for 2028 or 2029 might not have even started construction yet.
16:04
It’s not just Amazon’s money that has strings attached.
16:07
OpenAI’s exclusive deal with Microsoft only runs until 2032,
16:11
and there's reportedly a clause that voids parts of the whole agreement if OpenAI's
16:16
board formally declares it's achieved AGI. Two of the biggest checks OpenAI's ever gotten are both,
16:22
in some way, bets on the exact same unanswered scientific question.
16:26
When OpenAI first took outside money, it set itself up as a “capped-profit”
16:31
company. Investors could make money, sure, but only up to 100 times what they put in.
16:35
Anything beyond that flowed back to the nonprofit mission the company was
16:39
founded on. This was the thing OpenAI always pointed to when people asked how a company
16:44
that supposedly cared about user safety justified taking billions from Microsoft.
16:49
Years before anyone noticed, that cap got modified to allow a 20% annual increase. By the time
16:56
OpenAI officially converted into a for-profit public benefit corporation in October 2025,
17:01
that cap had basically been meaningless for years. The move just made it official.
17:06
There would be no ceiling on investor returns anymore.
17:09
This is what made going public possible to begin with.
17:12
OpenAI confidently filed its IPO paperwork with the SEC in May 2026, aiming for a public listing
17:19
as early as Q4 of that year. They want a valuation anywhere from $852 billion to $1 trillion.
17:26
It’s safe to say that not everyone who bought into the AI revolution is convinced it's paying off..
17:32
Klarna replaced 700 support agents with an OpenAI-built assistant, celebrated it, then
17:38
watched satisfaction scores drop before posting a loss. They quickly began rehiring humans again.
17:44
Their AI wasn't bad at answering questions. But it was definitely bad at everything else.
17:50
If the company OpenAI held up as proof of concept is already walking it back,
17:54
that might be a preview of what’s coming.
17:56
In the end, OpenAI is betting on 5 things landing at once.
18:01
That funding shows up on schedule; that its depreciation math on its GPUs holds up;
18:05
that its circular financing stays circular; that their much-vaunted ad market materializes;
18:11
and that cheap electricity stays cheap indefinitely.
18:15
None of these have to break simultaneously for the whole thing to collapse. If one slips,
18:20
suddenly the next round of funding costs more to raise. A company
18:23
balancing 5 bets at once loses the market’s confidence on all of them.
18:28
The AI bubble and the Dot-Com crash might be separated by 2 decades,
18:32
but the warning signs look disturbingly familiar. Find out what we might be
18:36
about ot face in “AI Bubble vs Dot Com Crash. History is REPEATING“. Or click on this video.
