Latest / The AI Practitioner Brief / Anthropic's Loss, DeepSeek's Profit, and Meta's Cluster — October 21, 2026

Transcript

Anthropic's Loss, DeepSeek's Profit, and Meta's Cluster — October 21, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_00: the podcast from AIPractitioner.substack.com,
• 0:10
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:13
  SPEAKER_00: for AI practitioners.
• 0:15
  SPEAKER_00: I'm Lina Faik,
• 0:16
  SPEAKER_00: an ML engineer,
• 0:17
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:21
  SPEAKER_00: It's Wednesday, October 21.
• 0:23
  SPEAKER_01: Morning.
• 0:23
  SPEAKER_01: Let's get into it.
• 0:24
  SPEAKER_00: With me is Jimmy.
• 0:26
  SPEAKER_00: My co-host on the brief, who asks the questions I skip past.
• 0:30
  SPEAKER_00: Anthropics leaked IPO paperwork shows an $80 billion loss in just six months.
• 0:35
  SPEAKER_00: DeepSeek says its own inference system is already profitable,
• 0:38
  SPEAKER_00: making $38 for every hour it runs a single chip.
• 0:42
  SPEAKER_00: And Meta is building a supercluster meant to reach 1 million of those chips.
• 0:46
  SPEAKER_00: That's the same industry running on wildly different math.
• 0:50
  SPEAKER_01: Those numbers don't even belong in the same paragraph.
• 0:52
  SPEAKER_00: We'll start with the headlines, then three deep dives.
• 0:55
  SPEAKER_00: Anthropik's enormous IPO numbers come first.
• 0:58
  SPEAKER_00: DeepSea claims it's already turning a profit on every chip it runs and will end with
• 1:02
  SPEAKER_00: the physical supercluster Meta is building to fuel all of it.
• 1:06
  SPEAKER_01: Alright, what's new today?
• 1:08
  SPEAKER_00: Google and Meta released an open-source project called Torch TPU this week.
• 1:13
  SPEAKER_00: It lets engineers run PyTorch,
• 1:15
  SPEAKER_00: the most widely used framework for building AI models,
• 1:18
  SPEAKER_00: directly on Google's TPUs,
• 1:20
  SPEAKER_00: which are Google's own AI chips,
• 1:22
  SPEAKER_00: instead of needing NVIDIA's GPUs, the chips most AI models currently run on.
• 1:27
  SPEAKER_01: Does it handle training or just running the model afterward?
• 1:30
  SPEAKER_00: Both.
• 1:30
  SPEAKER_00: And it's free to use,
• 1:32
  SPEAKER_00: which matters if a team is trying to cut how dependent it is on NVIDIA's pricing.
• 1:36
  SPEAKER_01: A real second option, not just a workaround.
• 1:39
  SPEAKER_00: Two other companies showed off homemade AI infrastructure this week, too.
• 1:43
  SPEAKER_00: Ramp, the corporate card company,
• 1:45
  SPEAKER_00: built its own framework for running AI agents,
• 1:48
  SPEAKER_00: software that can act across several steps on its own.
• 1:51
  SPEAKER_00: Ramp calls it laminar and says the whole thing was written by AI models over six weeks.
• 1:56
  SPEAKER_01: Six weeks for something handling production traffic?
• 1:59
  SPEAKER_00: That's Ramp's own claim.
• 2:00
  SPEAKER_00: Stack Overflow made a similar bet.
• 2:02
  SPEAKER_00: It launched a version of its site built for AI agents to query instead of people
• 2:06
  SPEAKER_00: and says a chat GPT plugin tripled its overnight signups.
• 2:10
  SPEAKER_01: So Stack Overflow's new customer is the agent, not the developer.
• 2:14
  SPEAKER_00: That's the AI side.
• 2:15
  SPEAKER_00: Elsewhere in tech today,
• 2:17
  SPEAKER_00: Elon Musk says Tesla and SpaceX have formed committees to discuss a possible merger.
• 2:22
  SPEAKER_00: Nothing is agreed yet and any deal would need a shareholder vote and regulatory approval.
• 2:26
  SPEAKER_01: That's a lot of paperwork for an idea with no agreement behind it.
• 2:30
  SPEAKER_00: PHP's maintainers are also warning admins about two critical vulnerabilities
• 2:34
  SPEAKER_00: that allow remote code execution across old and current PHP versions.
• 2:39
  SPEAKER_00: Emergency patches are coming.
• 2:40
  SPEAKER_01: That's most of the web's backend in one sentence.
• 2:43
  SPEAKER_00: Pretty much.
• 2:44
  SPEAKER_00: So this is a patch this week item.
• 2:46
  SPEAKER_00: Cloudflare also rewrote the TypeScript compiler in Rust this week,
• 2:49
  SPEAKER_00: working with a tooling company called Void Zero, and named the result TSRS.
• 2:54
  SPEAKER_00: Early benchmarks show it running up to eight times faster and using about 70% less memory.
• 3:00
  SPEAKER_01: Does it replace the existing compiler or sit next to it?
• 3:03
  SPEAKER_00: Still early,
• 3:04
  SPEAKER_00: but the real point is it can run inside WebAssembly,
• 3:07
  SPEAKER_00: a format for running code directly inside a browser.
• 3:10
  SPEAKER_00: That opens it up to lightweight environments the original compiler never reached.
• 3:14
  SPEAKER_01: Let's get into the big one then, the Anthropic numbers.
• 3:17
  SPEAKER_00: Why would a draft like this leak twice?
• 3:19
  SPEAKER_00: Reporting doesn't say who leaked it,
• 3:20
  SPEAKER_00: but a document this financially sensitive,
• 3:22
  SPEAKER_00: ahead of a listing this size,
• 3:23
  SPEAKER_00: is an obvious target for early leaks before the official filing.
• 3:42
  SPEAKER_01: What does it actually show?
• 3:43
  SPEAKER_00: $16.3 billion in revenue for the first half of 2026.
• 3:48
  SPEAKER_00: Against that, an $80.7 billion net loss in the same six months.
• 3:52
  SPEAKER_01: A loss five times the revenue.
• 3:54
  SPEAKER_01: How does a company spend that much?
• 3:56
  SPEAKER_00: Compute.
• 3:57
  SPEAKER_00: The filing says Anthropic spent $9 billion on compute and infrastructure in just that half year.
• 4:03
  SPEAKER_00: That's more than it spent in all of 2025 combined.
• 4:06
  SPEAKER_01: And they're planning to spend even more going forward?
• 4:09
  SPEAKER_00: Much more.
• 4:09
  SPEAKER_00: The filing says Anthropic could commit up to $750 billion to compute capacity over
• 4:15
  SPEAKER_00: the coming years.
• 4:16
  SPEAKER_00: That includes non-cancellable cloud deals with Google, Amazon, and Microsoft.
• 4:20
  SPEAKER_01: Non-cancellable, meaning locked in no matter what happens to revenue.
• 4:24
  SPEAKER_00: That's exactly the risk the filing itself names.
• 4:27
  SPEAKER_00: One outlet's read of the document says a large piece of that $80 billion loss,
• 4:32
  SPEAKER_00: something like $72 billion,
• 4:34
  SPEAKER_00: is a non-cash accounting charge tied to financial instruments Anthropic issued
• 4:38
  SPEAKER_00: to raise money.
• 4:39
  SPEAKER_00: Not cash actually leaving the building.
• 4:41
  SPEAKER_01: So the real cash burn might be smaller than the headline number.
• 4:44
  SPEAKER_00: Possibly,
• 4:45
  SPEAKER_00: but Anthropic hasn't confirmed that breakdown itself,
• 4:48
  SPEAKER_00: and the leaked draft isn't the final filing.
• 4:50
  SPEAKER_00: What is confirmed is the growth rate.
• 4:52
  SPEAKER_00: Revenue went from $4.6 billion for all of 2025 to $16.3 billion in just
• 4:59
  SPEAKER_00: the first half of this year.
• 5:00
  SPEAKER_01: That's genuinely fast.
• 5:01
  SPEAKER_01: Why go public now instead of staying private?
• 5:04
  SPEAKER_00: Because a $750 billion compute bill needs a bigger pool of money
• 5:08
  SPEAKER_00: than private funding rounds can keep supplying.
• 5:11
  SPEAKER_00: Anthropic is reportedly targeting a listing as soon as November,
• 5:14
  SPEAKER_00: at a valuation near $2 trillion.
• 5:17
  SPEAKER_01: A $2 trillion price tag on a company losing $80 billion in six months.
• 5:21
  SPEAKER_01: I don't see how public markets are supposed to price that kind of risk.
• 5:25
  SPEAKER_00: That's the bet being asked of investors, that revenue keeps compounding faster than losses do.
• 5:30
  SPEAKER_00: And this isn't unique to Anthropic,
• 5:32
  SPEAKER_00: other frontier labs are reportedly running similar math,
• 5:35
  SPEAKER_00: heavy losses next to fast growth,
• 5:37
  SPEAKER_00: because training and serving these models still costs more than anyone is charging
• 5:41
  SPEAKER_00: for them.
• 5:41
  SPEAKER_01: What's actually still unknown here?
• 5:43
  SPEAKER_00: Two things.
• 5:44
  SPEAKER_00: Whether that non-cash breakdown holds up once the real S1 is filed,
• 5:48
  SPEAKER_00: since leaked drafts get revised before anything becomes official.
• 5:52
  SPEAKER_00: And whether those cloud commitments survive if growth slows.
• 5:55
  SPEAKER_00: The filing itself lists preventing AI harm and securing enough compute
• 5:59
  SPEAKER_00: as named risk factors in Anthropic's own words.
• 6:02
  SPEAKER_01: So the risks they're flagging and the risks an investor should actually worry
• 6:05
  SPEAKER_01: about aren't quite the same list.
• 6:07
  SPEAKER_00: That's the split.
• 6:08
  SPEAKER_00: Investors are pricing growth,
• 6:10
  SPEAKER_00: the filing is flagging safety and supply risk,
• 6:13
  SPEAKER_00: and it's worth setting that $80 billion loss against a very different number
• 6:17
  SPEAKER_00: that came out of China this same week.
• 6:19
  SPEAKER_01: The company that's supposedly making money on this.
• 6:22
  SPEAKER_00: DeepSeq,
• 6:22
  SPEAKER_00: the Chinese AI lab known for undercutting Western pricing,
• 6:26
  SPEAKER_00: open-sourced something called DeepInfer this week.
• 6:28
  SPEAKER_00: It's the large-scale system DeepSeq uses to run inference,
• 6:32
  SPEAKER_00: meaning the step where a trained model actually answers a request,
• 6:35
  SPEAKER_00: for its own public service.
• 6:37
  SPEAKER_01: What does the system actually do?
• 6:39
  SPEAKER_00: DeepSeq says a single cluster now handles 155 trillion tokens a day.
• 6:44
  SPEAKER_00: A token is roughly a word or part of a word, the unit a model processes text in.
• 6:48
  SPEAKER_01: Trillion with a T. What does that mean for cost?
• 6:51
  SPEAKER_00: DeepSeq says that cluster generates $38 in revenue per chip per hour, based on off-peak pricing.
• 6:58
  SPEAKER_00: Most companies running inference at scale are still losing money per chip hour, not making it.
• 7:03
  SPEAKER_01: How does serving models even get to that point?
• 7:05
  SPEAKER_01: What's the actual trick?
• 7:07
  SPEAKER_00: A big part of it is something called KV Cache Compression.
• 7:10
  SPEAKER_00: When a model generates text,
• 7:11
  SPEAKER_00: it keeps a running memory of everything it's read so far,
• 7:14
  SPEAKER_00: called the key value cache.
• 7:16
  SPEAKER_01: And that memory is the expensive part?
• 7:18
  SPEAKER_00: Right.
• 7:19
  SPEAKER_00: That memory normally grows huge and costly to hold in chip memory.
• 7:23
  SPEAKER_00: DeepSeq's newest model,
• 7:24
  SPEAKER_00: nicknamed Flash,
• 7:25
  SPEAKER_00: compresses that cache aggressively,
• 7:27
  SPEAKER_00: so far more requests fit on the same hardware at once.
• 7:30
  SPEAKER_01: So the saving isn't a cheaper chip, it's squeezing more requests onto the same chip.
• 7:35
  SPEAKER_00: Exactly.
• 7:36
  SPEAKER_00: And DeepSeek built DeepInfer to run across different hardware, not just NVIDIA's.
• 7:40
  SPEAKER_00: Reporting on the release describes Huawei's chips getting,
• 7:43
  SPEAKER_00: in effect, equal billing with NVIDIA's in how the system is designed.
• 7:47
  SPEAKER_01: Is that confirmed or still a vendor claim?
• 7:49
  SPEAKER_00: The throughput and revenue figures come from DeepSeek's own disclosure,
• 7:53
  SPEAKER_00: so they're the company's numbers, not an independently run result.
• 7:56
  SPEAKER_00: Nobody outside DeepSeek has published a cluster at that scale to compare it against.
• 8:01
  SPEAKER_01: That's still a real number though, not a projection.
• 8:03
  SPEAKER_00: Right.
• 8:04
  SPEAKER_00: It's a reported result from a live system, not a future target.
• 8:07
  SPEAKER_00: What's unknown is whether it holds up outside DeepSeek's own infrastructure,
• 8:11
  SPEAKER_00: with different traffic or a smaller cluster.
• 8:14
  SPEAKER_01: What does a team actually do with that number?
• 8:16
  SPEAKER_00: Use it as a benchmark.
• 8:18
  SPEAKER_00: If your chip hours are costing you money instead of making $38 an hour,
• 8:22
  SPEAKER_00: the gap is probably sitting in how your cache and batching are set up,
• 8:25
  SPEAKER_00: not in which chip you're renting.
• 8:27
  SPEAKER_01: That's a very different story from the one we just heard about Anthropic.
• 8:31
  SPEAKER_00: It is.
• 8:31
  SPEAKER_00: Anthropic is burning money to build the next frontier model.
• 8:35
  SPEAKER_00: DeepSeek is squeezing profit out of serving an existing one.
• 8:38
  SPEAKER_00: Both of those need somewhere to physically run, which is where Meta comes in.
• 8:42
  SPEAKER_01: The supercluster you mentioned earlier.
• 8:44
  SPEAKER_00: Meta published details this week on Prometheus,
• 8:46
  SPEAKER_00: the supercluster it's building in New Albany, Ohio.
• 8:49
  SPEAKER_00: The site already runs hundreds of thousands of Nvidia chips,
• 8:52
  SPEAKER_00: and Meta says it's scaling toward one million.
• 8:55
  SPEAKER_01: A million chips in one place, what does that even run on power-wise?
• 8:59
  SPEAKER_00: Meta calls it gigawatt scale,
• 9:01
  SPEAKER_00: meaning the site is built to draw roughly as much electricity as
• 9:04
  SPEAKER_00: a small country's grid.
• 9:05
  SPEAKER_00: The first phase already trained Muse, Meta's family of AI companion models.
• 9:10
  SPEAKER_01: How does that compare to other AI data centers being built right now?
• 9:13
  SPEAKER_00: It's in the same league as XAI's Colossus cluster or OpenAI's Stargate project,
• 9:18
  SPEAKER_00: all racing towards similar scale around the same time.
• 9:21
  SPEAKER_00: The difference is Meta is one of the few publishing real reliability numbers instead
• 9:26
  SPEAKER_00: of just a headline chip count.
• 9:27
  SPEAKER_01: Is a million actually running today or is that the target?
• 9:31
  SPEAKER_00: Target, not current count.
• 9:33
  SPEAKER_00: Today it's two platforms.
• 9:34
  SPEAKER_00: One, called Catalina, runs on NVIDIA's GB200 chips.
• 9:38
  SPEAKER_00: The other, called Clemente, runs on the newer GB300.
• 9:41
  SPEAKER_01: Is that the full build, or is there more coming?
• 9:43
  SPEAKER_00: There's a third platform planned,
• 9:45
  SPEAKER_00: called San Miguel,
• 9:46
  SPEAKER_00: built on NVIDIA's upcoming Vera Rubin chips, the generation after GB300.
• 9:51
  SPEAKER_01: What's the actual engineering problem at that size?
• 9:54
  SPEAKER_01: Surely it's not just bolting on more racks.
• 9:57
  SPEAKER_00: It's keeping the thing running.
• 9:58
  SPEAKER_00: Meta says it cut training interruptions by 99%, down from a peak of a thousand a day.
• 10:04
  SPEAKER_00: At that scale, one failed chip or network link can stall the whole training run.
• 10:08
  SPEAKER_01: A thousand interruptions a day sounds like the system was barely usable before the fix.
• 10:13
  SPEAKER_00: It was close to that.
• 10:14
  SPEAKER_00: Meta measures this with something it calls good put.
• 10:17
  SPEAKER_00: The share of total compute time actually spent making training progress instead
• 10:20
  SPEAKER_00: of recovering from a failure.
• 10:22
  SPEAKER_00: It says the first phase hit over 85% good put.
• 10:25
  SPEAKER_01: I've also seen reporting that more than 11% of the cluster's chips are dead or held
• 10:30
  SPEAKER_01: in reserve at any given time.
• 10:32
  SPEAKER_00: That's a real number from outside reporting, not something meta itself has highlighted.
• 10:36
  SPEAKER_00: It's a reminder that even a well-run cluster this size loses a meaningful slice
• 10:41
  SPEAKER_00: of its hardware to failure and redundancy.
• 10:43
  SPEAKER_01: So the headline chip count is never the number actually doing useful work.
• 10:47
  SPEAKER_00: Right.
• 10:47
  SPEAKER_00: And that gap is exactly why goodput,
• 10:50
  SPEAKER_00: not raw chip count, is becoming the number series infrastructure teams watch.
• 10:54
  SPEAKER_01: What does that mean for a team that isn't building a meta-sized cluster?
• 10:57
  SPEAKER_00: The same math shows up at any scale, just with smaller numbers.
• 11:01
  SPEAKER_00: A team renting a few hundred chips loses a similar share to failures
• 11:05
  SPEAKER_00: and idle redundancy that Meta does at a million.
• 11:08
  SPEAKER_00: The difference is nobody publishes a good put number for a cluster that size,
• 11:12
  SPEAKER_00: so most teams never find out how much they're actually wasting.
• 11:15
  SPEAKER_01: That's a strange kind of invisible cost.
• 11:17
  SPEAKER_00: So that's the day.
• 11:18
  SPEAKER_00: Anthropic is burning 80 billion dollars to stay ahead on the next model,
• 11:22
  SPEAKER_00: DeepSeek is squeezing a profit out of serving an existing one,
• 11:26
  SPEAKER_00: and Meta is building the physical plant underneath all of it.
• 11:29
  SPEAKER_00: Three very different bets on the same race.
• 11:31
  SPEAKER_01: I'd bet the deep-infer number gets picked apart hardest of the three,
• 11:34
  SPEAKER_01: not because it's wrong,
• 11:36
  SPEAKER_01: but because it's the one other companies can actually measure themselves against.
• 11:40
  SPEAKER_00: That's Wednesday.
• 11:41
  [music]