Latest / Elon Musk Podcast / Anthropic Overtakes OpenAI in Annual Revenue

Transcript

Anthropic Overtakes OpenAI in Annual Revenue · Elon Musk Podcast

• 0:00
  Entropic has crossed a $30 billion annualized revenue run
• 0:03
  rate, officially overtaking Open AI as the highest grossing
• 0:07
  artificial intelligence company in the world.
• 0:10
  Yeah. And, and I mean, the math on
• 0:11
  this is just staggering when you actually put it in perspective,
• 0:14
  right? You have a company that you know
• 0:16
  barely existed a short while ago, and they are now generating
• 0:19
  double the combined annual revenue of Snowflake, Datadog,
• 0:24
  Cloudflare, Mongo DB and HubSpot.
• 0:27
  Oh wow. Yeah, you could add up the
• 0:29
  entire financial output of all those highly established,
• 0:32
  globally utilized software platforms and you would still
• 0:35
  fall roughly $15 billion short of where Anthropic sits today.
• 0:39
  That is just. I mean, it's hard to even
• 0:40
  process that scale. It is we are looking at this
• 0:43
  quiet takeover of the enterprise sector, propelled by, frankly,
• 0:47
  some highly unusual accounting mechanics, plus a
• 0:50
  record-breaking hardware order that totally redefines corporate
• 0:53
  scaling and a sudden clash with the Pentagon.
• 0:55
  So how does a company bypass the creator of the consumer
• 0:59
  artificial intelligence category without a viral app while
• 1:02
  spending a fraction of the money?
• 1:04
  And is that $30 billion number actually as solid as it looks?
• 1:08
  Well, Anthropic grew its revenue from 1 billion to $30 billion
• 1:13
  over a single 15 month period by basically entirely ignoring the
• 1:19
  everyday consumer. Right, which is wild.
• 1:21
  Yeah, they bypassed the consumer market completely and focus
• 1:24
  strictly on enterprise application programming
• 1:27
  interface contracts. They just went straight to the
• 1:29
  largest corporations on earth. Because if you look at Open AI,
• 1:32
  they have roughly 900 million weekly users interacting with
• 1:35
  their consumer products. Yeah, people are, you know,
• 1:38
  asking for recipes, summarizing articles, helping their kids
• 1:40
  with homework. Right, the everyday stuff.
• 1:42
  Exactly. Anthropic has a tiny fraction of
• 1:45
  that audience. I mean, their consumer footprint
• 1:47
  is negligible. We back up.
• 1:49
  Open AI has nearly a billion users, and Anthropic is beating
• 1:53
  them on revenue. Yeah, the monetization gap tells
• 1:55
  the real story here. Anthropic pulls in roughly 211
• 1:59
  dollars per monthly user. Open AI is extracting about $25
• 2:04
  per weekly user. Anthropic secured 8 of the
• 2:08
  Fortune 10 companies and they recently doubled their high
• 2:11
  value contracts. Wow.
• 2:13
  Yeah, they currently have over 1000 enterprise clients spending
• 2:16
  more than $1,000,000 annually. I mean securing $1,000,000
• 2:21
  annual contract with a Fortune 10 company that fundamentally
• 2:24
  changes the relationship between a software provider and a
• 2:27
  client. How so?
• 2:29
  Well, you aren't just selling them a separate tool that
• 2:33
  employees open in a web browser. You are weaving your underlying
• 2:37
  models directly into their daily operations.
• 2:40
  Right, it becomes infrastructure.
• 2:41
  Exactly. Think about how a global bank
• 2:44
  operates. If you are a customer of that
• 2:46
  bank and you upload a mortgage application, that application
• 2:49
  goes into a highly secure internal routing system, right?
• 2:53
  The bank integrates entropics model directly into that routing
• 2:56
  system to instantly read, classify, and extract the
• 2:59
  compliance data from your application.
• 3:01
  The model becomes a permanent piece of the plumbing.
• 3:03
  Which means the switching costs become astronomically high.
• 3:06
  Oh. Absolutely.
• 3:07
  Because if you are an individual consumer, you might, you know,
• 3:10
  cancel a $20 monthly subscription because you get
• 3:13
  bored or you just want to try a competitors app.
• 3:16
  Sure, happens all the time. But a global financial
• 3:19
  institution is not going to easily RIP out the core
• 3:22
  technology that processes all of their sensitive compliance
• 3:25
  documentation. Right, because doing so would
• 3:27
  require basically freezing their operations.
• 3:30
  Exactly. They'd have to rewrite thousands
• 3:32
  of lines of integration code and retrain their entire compliance
• 3:37
  department. Which is why this flips the
• 3:38
  assumed tech playbook totally on its head.
• 3:40
  Yeah, for the last two decades, the accepted wisdom in Silicon
• 3:44
  Valley was that consumer virality was a prerequisite for
• 3:47
  Absolute dominance. You needed a billion people to
• 3:51
  download your software. Right, and Fropic proves that
• 3:54
  consumer virality is entirely unnecessary.
• 3:57
  By targeting existing corporate workflows, you build durable,
• 4:01
  high value revenue that just compounds instantly.
• 4:04
  You just convinced the largest corporations on earth to route
• 4:07
  their existing internal software budgets through your servers.
• 4:10
  Exactly every single time a corporate client's internal
• 4:13
  software pings the Anthropic model to analyze, say, A50 page
• 4:17
  legal contract, that client pays for the exact amount of
• 4:21
  computing power used plus a margin.
• 4:24
  So it functions like a utility company charging for water.
• 4:27
  Yeah, exactly like a utility. Consumer models require massive,
• 4:31
  expensive infrastructure just to handle millions of casual
• 4:35
  queries from users on free tiers, which, you know,
• 4:38
  generates very little direct. Revenue.
• 4:40
  So they just skipped the App Store, walked straight into the
• 4:42
  boardroom and started taking over software budgets.
• 4:44
  That's exactly it. They established a utility
• 4:47
  monopoly over corporate operations and a massive driver
• 4:51
  of this specific enterprise revenue is a tool they built
• 4:55
  called Claude Code. This single product went from
• 4:58
  non existent to a 2 1/2 billion dollar run rate in less than a
• 5:02
  year. 2 1/2 billion dollars for a single coding tool.
• 5:05
  Yeah. I mean, most successful software
• 5:08
  startups dream of hitting $100 million in total revenue before
• 5:12
  they file for an initial public offering.
• 5:13
  Oh, for sure. And Claude code hit 25 times
• 5:17
  that amount in a matter of months.
• 5:19
  The mechanics of how Claude code integrates into the software
• 5:22
  development process explain that financial velocity.
• 5:25
  Really. It currently writes 4% of all
• 5:28
  public GitHub commits on Earth, and that number is projected to
• 5:32
  hit 20% shortly, which is huge. It is To understand the gravity
• 5:36
  of that we have to look at what a commit actually is.
• 5:40
  A commit is essentially a saved change or a new addition to a
• 5:44
  softer project source code. It is the fundamental building
• 5:47
  block of software development. So whenever an engineer fixes a
• 5:50
  bug or adds a new feature, they submit a commit, Yes.
• 5:54
  And 4% of all those updates globally are now authored
• 5:58
  entirely by this system. It is like having an invisible
• 6:02
  senior level developer permanently attached to every
• 6:04
  programmer's keyboard, continuously writing and
• 6:07
  checking their work. That's a great way to put it.
• 6:09
  If you are a programmer, you are no longer just staring at a text
• 6:12
  editor trying to remember the exact syntax for a complex
• 6:15
  database query. Right, you have this autonomous
• 6:18
  partner constantly analyzing the entire structure of your
• 6:21
  project. Exactly.
• 6:23
  It drafts entire blocks of logic, spots security
• 6:26
  vulnerabilities you missed, and updates legacy systems
• 6:29
  concurrently while you work and other things.
• 6:31
  And we can see exactly how this alters the traditional workflow
• 6:34
  by looking at the compliance security company Drata.
• 6:36
  Right. The Drata examle is fascinating.
• 6:39
  They had a team of over 80 engineers.
• 6:42
  Traditionally, software engineering requires a massive
• 6:46
  amount of manual quality assurance.
• 6:48
  Oh, a ton of it. Like if you write a piece of
• 6:49
  code to handle user passwords, you then have to write 10 other
• 6:53
  pieces of code just to test if the first piece fails under
• 6:56
  weird conditions. Like if a user types in entirely
• 6:59
  special characters, or if the server connection drops halfway
• 7:01
  through. Right, it is incredibly tedious
• 7:04
  but necessary work. Andrada used this technology to
• 7:07
  basically eliminate that manual testing phase.
• 7:10
  They achieved 86% faster quality assurance cycles and executed 4
• 7:15
  times as many test cases. Which totally changes the core
• 7:18
  function of a software engineering department, Yeah.
• 7:20
  It really does. If an AI agent can write and
• 7:22
  execute four times the number of test cases in a fraction of the
• 7:26
  time, the human engineers stop being writers of code and start
• 7:30
  becoming managers of artificial intelligence output.
• 7:33
  Well, I do want to push back on the hype a little bit here
• 7:35
  though. OK, go ahead.
• 7:36
  We are looking at explosive growth, but companies currently
• 7:41
  have massive dedicated budgets specifically earmarked for
• 7:46
  artificial intelligence experimentation.
• 7:49
  They are paying premium prices to automate the straightforward
• 7:52
  tasks like writing basic test cases or generating boilerplate
• 7:56
  code. But once all the easy,
• 7:58
  repetitive tasks are automated across these enterprise clients,
• 8:02
  that explosive revenue growth could easily plateau.
• 8:05
  You think it's a bubble? I'm just saying a 2 1/2 billion
• 8:08
  dollar run rate is incredible, but it could represent a rush to
• 8:11
  grab the lowest hanging fruit and corporate IT departments
• 8:14
  rather than, you know, a perpetual growth.
• 8:16
  Engine, I see what you mean, but the efficiency gains point to a
• 8:19
  permanent structural shift in human labor rather than just a
• 8:22
  temporary IT budget trend. You really think so?
• 8:25
  I do. When a company realizes they can
• 8:27
  deploy software 86% faster, the financial incentive to maintain
• 8:32
  that speed permanently alters their business model.
• 8:35
  Well, yeah, the competitive pressure dictates that any
• 8:37
  software company not utilizing these autonomous coding agents
• 8:40
  will just be outpaced by those that do.
• 8:42
  Exactly. Furthermore, this severely
• 8:45
  shrinks the hiring pipeline for junior developers.
• 8:48
  Entirely. Historically, junior developers
• 8:51
  cut their teeth and learn the architecture of a company by
• 8:53
  writing test cases and doing that basic quality assurance
• 8:57
  right. If an invisible developer
• 8:59
  handles that instantly, the entry level rung of the industry
• 9:02
  ladder vanishes. You are raising the baseline
• 9:05
  speed of software deployment globally, which requires
• 9:08
  continuous permanent API usage. Which naturally leads to the
• 9:12
  cost of maintaining that continuous usage, right?
• 9:15
  Anthropic's reported model training costs are surprisingly
• 9:18
  low. Actually they spend roughly 1/4
• 9:21
  of what open AI spends to train their models.
• 9:23
  Wow, just 1/4. Yeah, but they're $30 billion
• 9:27
  revenue figure is calculated using an accounting method that
• 9:30
  requires some pretty careful examination.
• 9:33
  OK, let's talk about the discrepancy there.
• 9:35
  Anthropic reports their revenue on a gross basis.
• 9:39
  This means they include the percentage cut that their cloud
• 9:42
  partners, specifically Amazon and Google, take from the sales.
• 9:46
  Open AI, conversely, reports net revenue.
• 9:49
  They subtract the roughly 20% share they pay to Microsoft
• 9:53
  before they report their final revenue number.
• 9:55
  Hold on, if Anthropic is counting Amazon's cut as their
• 9:58
  own revenue, is this $30 billion number actually real?
• 10:02
  Let me get this straight. Sure.
• 10:03
  Yeah, if Anthropic sells $10.00 of access to their API through
• 10:07
  Amazon's cloud infrastructure, and Amazon's contract entitles
• 10:11
  them to keep $3 of that sale, Anthropic still reports the full
• 10:15
  $10 as their top line revenue. Exactly.
• 10:17
  They reported the $10 as revenue and then they record the $3 as
• 10:21
  an operating expense. Open AI in that exact same
• 10:24
  scenario with Microsoft, would simply report $7.00 of revenue.
• 10:27
  Generally accepted accounting principles allow for both
• 10:30
  methods, depending on how the contracts are structured and who
• 10:33
  holds the primary responsibility for delivering the service.
• 10:37
  But it fundamentally changes how you compare the financial
• 10:39
  velocity of the two companies. You have to make a significant
• 10:43
  mental adjustment to the $30 billion figure.
• 10:46
  Definitely. The deeper mechanical issue here
• 10:48
  is the compute for equity dynamic.
• 10:51
  Yeah, this is where the financial architecture gets
• 10:53
  incredibly circular. Well, very much so.
• 10:56
  Amazon and Google invested 10s of billions of dollars into
• 10:59
  Anthropic. They provided massive infusions
• 11:02
  of venture capital. But entropics does not just sit
• 11:06
  on that cash. They take that investment
• 11:08
  capital and spend it heavily right back on the exact cloud
• 11:11
  service servers owned by Amazon and Google.
• 11:14
  Yeah, in one recent period, Anthropic spent over 2 1/2
• 11:16
  billion dollars on Amazon Web Services.
• 11:20
  Against that massive expense, they generated roughly the same
• 11:23
  amount of revenue through Amazon.
• 11:25
  If we think about this mechanically, it is remarkably
• 11:28
  similar to a casino offering a High Roller a complimentary
• 11:32
  penthouse suite, provided the High Roller agrees to gamble the
• 11:36
  exact cash value of that suite at the blackjack tables
• 11:39
  downstairs. That's a perfect analogy.
• 11:41
  The casino can report high occupancy rates and massive
• 11:44
  activity on the casino floor, but no new external cash has
• 11:47
  actually entered the ecosystem. Amazon invests billions in
• 11:51
  Anthropic, Anthropic pays those billions back to Amazon for
• 11:55
  computing power, and Amazon sells Anthropic's technology to
• 11:58
  corporate clients. The money moves in a tight
• 12:01
  closed loop. And this accounting choice and
• 12:03
  the closed loop spending directly impact how the
• 12:06
  secondary venture capital market values the companies, right
• 12:10
  Right now, investor demand for Anthropic shares is intense.
• 12:13
  Goldman Sachs charges a premium carry, which is an additional
• 12:16
  percentage of the profits taken by the firm, simply to give
• 12:19
  clients access to Anthropic shares.
• 12:22
  Meanwhile, millions of dollars in open AI shares sit unsold on
• 12:25
  secondary markets. The gross revenue reporting
• 12:28
  method creates a perception of sheer financial momentum that
• 12:31
  investors find irresistible. But this opens up severe
• 12:34
  regulatory scrutiny for when Anthropic eventually files for
• 12:38
  an initial public offering. Public market auditors and the
• 12:42
  Securities and Exchange Commission will dissect that
• 12:45
  gross revenue figure meticulously.
• 12:48
  They will demand absolute transparency regarding exactly
• 12:51
  how much of that $30 billion is flowing freely into Anthropic's
• 12:55
  bank accounts, clear of cloud provider fees and circular
• 12:59
  spending arrangements. Because the private venture
• 13:02
  market thrives on momentum and top line growth metrics, but the
• 13:05
  public equities market requires rigorous standardization and
• 13:08
  cash flow transparency. Exactly.
• 13:10
  And to justify those secondary market valuations to future
• 13:13
  public auditors, Anthropic has to prove they can physically
• 13:17
  scale the utility of these models.
• 13:19
  Right. Which brings us to the recent
• 13:20
  agreement they signed with Google and Broadcom for 3 1/2
• 13:24
  gigawatts of TPU compute capacity.
• 13:26
  Let us pause to comprehend the physical reality of 3 1/2
• 13:30
  gigawatts. It's massive.
• 13:31
  We are talking about an order for nearly 1,000,000 custom
• 13:35
  processor chips. Those chips will consume enough
• 13:39
  raw electricity to power an American city of 700,000 people.
• 13:43
  It is. Wild.
• 13:44
  And this massive power allocation is solely for running
• 13:47
  the current models, right? It is entirely for processing
• 13:51
  the enterprise API requests, routing the compliance
• 13:55
  documents, and writing the GitHub commits.
• 13:58
  It does not even factor in the energy required to train their
• 14:01
  future, more advanced models. We are measuring tech companies
• 14:04
  by employee headcount anymore. No, we are measuring them by raw
• 14:08
  electricity consumption. If you look back at historical
• 14:11
  industrial booms, scaling an empire was always tied to
• 14:14
  physical constraints. Standard Oil built their
• 14:17
  monopoly through the physical construction of miles of iron
• 14:20
  pipelines across the country. They had to secure the land, lay
• 14:23
  the pipe and pump the crude. General Motors required over
• 14:26
  400,000 employees to hit its production peak during wartime
• 14:30
  manufacturing. They needed vast factory floors,
• 14:33
  millions of tons of steel, and massive human coordination.
• 14:37
  Enthropic is generating an equivalent financial empire
• 14:40
  without the iron pipes or the assembly line workers.
• 14:44
  The primary raw material of their expansion is electricity
• 14:47
  flowing through silicon. And this shifts the ultimate
• 14:51
  bottleneck of the entire intelligence economy.
• 14:54
  Yeah, growth is no longer limited by how many software
• 14:57
  sales representatives they can hire or how many top tier
• 15:01
  engineers they can recruit from universities, right?
• 15:03
  The growth is strictly limited by the physical constraints of
• 15:07
  regional power grids and semiconductor manufacturing
• 15:10
  capacities. Because you cannot negotiate
• 15:12
  with a power grid exactly If there is no electricity
• 15:15
  available in a specific geographic region, you simply
• 15:18
  cannot deploy the million custom chips.
• 15:21
  Negotiating for 3 1/2 gigawatts is not a software problem.
• 15:24
  It is a heavy infrastructure problem, Absolutely.
• 15:27
  It requires sitting down with local utility monopolies,
• 15:30
  securing land permits from municipal governments, and
• 15:33
  physically building substations and cooling facilities years in
• 15:36
  advance. And securing that power is a
• 15:39
  profound competitive advantage. Any competitor trying to catch
• 15:43
  up to Anthropic does not just need to write better algorithms,
• 15:47
  right? They need to somehow locate
• 15:49
  unused power plants and secure transmission lines that have not
• 15:53
  already been claimed. And that deep reliance on
• 15:56
  physical infrastructure and government cooperation ties
• 16:00
  directly into a recent development with the United
• 16:02
  States Department of Defense. Yeah, this is fascinating.
• 16:05
  The Pentagon recently designated Anthropic a supply chain risk
• 16:10
  after the company refused to compromise on its acceptable use
• 16:13
  policy. Which is a highly unusual
• 16:15
  situation for a major technology contractor.
• 16:17
  Very unusual. Traditionally, companies will
• 16:19
  bend over back to secure lucrative defense contracts
• 16:22
  because the revenue is incredibly stable and massive in
• 16:25
  scale, but Anthropic drew hard, non negotiable lines regarding
• 16:30
  how their models could be utilized by military personnel.
• 16:33
  Specifically, they drew hard lines stating their models
• 16:36
  cannot be used for the mass domestic surveillance of
• 16:38
  citizens, nor can they be used to power fully autonomous
• 16:42
  weapons systems, meaning systems that can select and fire on
• 16:47
  human targets without any human intervention or final approval.
• 16:51
  But the Pentagon demanded unrestricted access, insisting
• 16:55
  on the ability to use the technology for any lawful
• 16:58
  purpose. The mechanical conflict here
• 17:00
  centers entirely on the definition of control and
• 17:03
  liability in defense contracting.
• 17:05
  Explain that. Well, the Pentagon operates on
• 17:07
  the principle that a private civilian contractor should not
• 17:10
  dictate the operational parameters of the military,
• 17:12
  provided the military is acting within the bounds of national
• 17:15
  and international law. Anthropic operates on the
• 17:19
  principle that the potential consequences of autonomous
• 17:22
  intelligence and warfare are too severe to leave open to the
• 17:25
  interpretation of a military commander in the field.
• 17:28
  And this friction has sparked intense public debate.
• 17:31
  Oh, definitely. We should impartially note that
• 17:33
  Elon Musk publicly criticized Anthropic over this specific
• 17:36
  decision. Yeah, he did.
• 17:37
  He labeled the company as misanthropic and evil.
• 17:41
  He also claimed their models are discriminatory against whites,
• 17:45
  Asians, heterosexuals and men. And to be clear, we are simply
• 17:49
  reporting these statements as part of the broader public
• 17:52
  reaction that's in the sources. We aren't taking a position on
• 17:55
  them. Absolutely.
• 17:56
  But the critical insight here is how this ethical friction
• 18:01
  completely alters the market dynamics of government
• 18:04
  procurement. Because Anthropic locked
• 18:06
  themselves out of certain military contracts to maintain
• 18:09
  their strict internal policies, they are inadvertently acting as
• 18:13
  a massive market subsidy for smaller, unregulated startups.
• 18:17
  Because the Pentagon still needs the technology, right?
• 18:20
  If the $30 billion giant refuses the contract over the any lawful
• 18:24
  purpose clause, a smaller artificial intelligence startup
• 18:27
  will gladly agree to those terms.
• 18:29
  Exactly. A startup with a fraction of the
• 18:31
  computing power suddenly secures a massive defense agreement
• 18:34
  simply because they do not have the same internal restrictions
• 18:37
  regarding autonomous weapons systems.
• 18:39
  Which means Anthropic stance guarantees that military
• 18:42
  applications will be developed and deployed by secondary
• 18:45
  companies, companies that operate with far less public
• 18:49
  visibility and potentially fewer internal safeguards.
• 18:53
  It creates a completely separate shadow ecosystem of defense
• 18:57
  focused artificial intelligence companies building the exact
• 19:01
  tools Anthropic refused to build.
• 19:03
  So, anthropic bypass consumer trends to quietly dominate the
• 19:07
  enterprise sector, generating more revenue faster than any
• 19:11
  corporation in history, while simultaneously challenging the
• 19:14
  military over the ethical boundaries of automated
• 19:17
  intelligence. What happens to the global
• 19:19
  economy when a single private company controls the invisible
• 19:22
  infrastructure, writing 1/5 of the world software, and their
• 19:26
  only real limitation is finding enough electricity to keep the
• 19:29
  physical servers running? If you're not subscribed yet,
• 19:32
  take a second and hit follow on whatever app you're using.
• 19:34
  It helps us keep making this. We appreciate you being here.