the Polo Grounds

Back to the forum index

Linear

This is the next chapter of AI race -- firms are moving away from venture backed and into public markets for funding.

The S1 will be extremely revealing in terms of the real economics of AI. Things like revenue growth, margins, data center commits, compute expenses, etc are all massive speculation points today. We're going to find out really quickly whether these frontier model companies are amazing businesses or cash furnaces.

The thing I'm watching most closely isn't the IPO itself. It's whether the eventual filing shows that Anthropic has discovered a path to sustained margins despite the enormous compute costs.

--
Freeman's got so much riz they need to call him Aura Parseghian

352 views

A student in the US is using an agent for coding and can select the AI he uses. At first he let it ride with the US ones, quickly burned through $200 and stopped. He switched to letting the agent use DeepSeek, and has done far more without paying much at all. I had heard that DeepSeek was 10% the cost, but from his use, seemed closer to 1%.

47 views

Are there any other companies that have as many corporate contracts as Anthropic does?

I know Salesforce started using Anthropic a few years ago.

Everything I read about the other options in this space just makes me think Anthropic is the clear choice (in a market sector Im not 100% sold on).

I assume Anthropic’s IPO is partially to keep up with the jones’ bc they dont want their competition to catch up to them by getting a ton of money from the public via IPOs.

240 views

What I struggle with is understanding how any of the frontier model companies become sustainably profitable businesses. Anthropic appears to have tremendous enterprise adoption and some of the strongest corporate relationships in the industry, but they're also operating in a market where compute consumption seems to rise as fast as demand. Reports around the IPO suggest revenue is growing explosively, but so are infrastructure commitments and capital requirements. They're reporting as "profitable" in Q1/Q2 of 2026 largely on "trust me bro" accounting principles and due to a massive computing discount from SpaceX that expires in Q3.

The question I keep coming back to is whether these companies eventually look like software businesses with software margins, or whether they end up looking more like utilities, cloud providers, or telecoms that require enormous ongoing capital investment.

If Anthropic can demonstrate a path to strong margins, the IPO could be a watershed moment. If not, then even being the market leader may not translate into the kind of economics investors are expecting.

--
Freeman's got so much riz they need to call him Aura Parseghian

179 views

I don't believe they will.

Current subsidy rates for LLM pricing are estimated at 80%. Even with that major discount, CIOs in large corporate are already putting the brakes on ummetered AI rollouts and rehiring junior developers. The scale projections for Anthropic et al. are still in the handwave stage of rigor. They won't reach the scale they think.

Smart players, IMO, have already figured this out. This is an environment where you sell pickaxes and let the gold miners take all the risk. The better-led firms are playing it that way. Microsoft very carefully hedged its bets on OpenAI all along. Apple, a firm not shy about spending R&D money, has let others take the lead and as usual will swoop in late. IBM pretty clearly is setting up its usual middleware play. X it's hard to tell what's going on but them renting compute time to Anthropic tells you something.

It will all end in tears. Remember Global Crossing?

97 views

I just don't know how big the blast radius is going to end up being. Unlike the post dot com internet expansion era, this isn't overbuying a bunch of infrastructure that's cheap to operate. Even if the Data Centers/GPUs were free, you'd go broke operating them.

--
Freeman's got so much riz they need to call him Aura Parseghian

98 views

The circular financing involved in AI model development <> compute and infrastructure has swamped actual valuations at this point.

If OpenAI's orders drop off the books for NVIDIA....

There's a reason OpenAI is pulling products and "focusing on core." And frantically getting their IPO out the door.

In my view, firms are going to need both model development capability AND model distribution/monetization capability to be successful.

OpenAI has the former, but not the latter.
Anthropic did a decent job on the latter by concentrating on corporate customers.
Oracle I can't figure out other than Ellison's had too much Michigan hookers and blow.
Microsoft has a hand in the former but is focused on the latter and figure if OpenAI goes away, there's many fish in the sea.
Mistral I've lost track of, no idea.

Winners/survivors:
Google
X
probably Anthropic

And to your point on data centers: not only do you have to operate them, you have to rebuild them every five years as compute power continues to scale. Not cost-justified for many smaller workloads.

99 views

I’m pretty sure there’s a ton of chips on boxes and my feeling is the SEC is going to not like how they’ve classified revenue the last 5 years.

--
Freeman's got so much riz they need to call him Aura Parseghian

92 views

But they've got a HUGE problem in that they don't have a platform for it, and haven't really made any progress on that front since it became relatively obvious that it was a problem.

Companies are finding great value from letting their top talent use it, and figure out how to use it collaboratively. But, that level of autonomy doesn't really work all that well in companies who want to have much more control over governance and permissions, and the amount of workarounds necessary to collaborate in usual workflows with Claude are substantial.

Honestly, I'd put my money on Microsoft. Their sharepoint platform more or less runs the corporate world, and copilot natively integrates. Copilot doesn't have to be as good as Claude, it just has to be good enough. Put another way, I think it's more likely that Microsoft develops a good enough AI product than it is that Anthropic develops a good enough collaborative corporate work platform.

I had been surprised by Google's seeming uninterest in this side of things, but I think they're making their play as being the personal AI product of choice, and thus have been building out for much different use cases.

GPT, after serving a short stint as the Myspace of AI, I think actually has a superior product to Claude in Codex right now for many use cases. Claude is still generically better for "office" work and tone, but Codex feels smarter to me again. Consider Claude the polished manager who can be client-facing, but Codex is the autistic engineer who can get shit done.

220 views

Ive never seen an AI completely ignore guardrails more than copilot.

Moreover, I‘ve seen examples of copilot straight up re-writing the guardrail because it didnt think it should have a guardrail.

Copilot is a serious risk that‘s not worth it imo.

59 views

- No text -

74 views

and it taps into all the main models (gpt, Claude, etc) to do various things. I started putting together an agent in Copilot Studio and found that you can actually select the model to use.

154 views

And at the time it was truly terrible. If they're just letting other products plug in as a connector they're setting up well to deploy their own good enough product later while getting people trained on using their harness using other people's product.

34 views

I like that firms are starting to specialize a bit, that should at least drive utility for specific applications.

It is going to be very interesting to see where price points end up. Everyone loves AI when it's free or cheap...

--
At night, the ice weasels come.

143 views

the "problem" is much narrower than stated. You can't just randomly replace wide swaths of employees with LLMs, or expect that substandard employees are going to magically turn in better work when provided LLM resources.

It's more like that it makes otherwise good employees more efficient. You still need a capable human in the loop.

134 views

It makes good people and good processes great. It makes lackluster people and poor process even more of a bottleneck and impediment to faster velocity.

121 views

My big picture view is that those who are good at using AI are good at managing employees. Those who have a clear vision of what it is they want and are trying to do will be able to be very effective with AI. Those who don't will struggle.

And those things are also true at the larger level. The penalty for bad processes will be higher now, as those with good processes will be so much more efficient and productive.

Management consultants rejoice.

26 views

It'll be really interesting when the OSS LLMs and the other 3 or 4 big players squeeze the hell out of each other from price pressure.

--
Sometimes I rhyme slow sometimes I rhyme quick.

235 views

On one hand, if they were was making money hand over fist, they would have no problem accessing all of the capital they need. So, what's the point of going public when you can keep all of those tasty profits for yourself?

On the other hand, it is quite possible that the AI frenzy means that demand for the stock will send valuations through the roof. So, even if they have a good business, it is likely worth more to the current investors to sell to the next greater fool, whether they are actually a fool or not.

It vaguely reminds me of a quote from our CEO during the dot com heyday: "Going public is not a business plan."

--
At night, the ice weasels come.

253 views

to watch Elon continue to pivot to focus on AI infrastructure. His ability to get datacenters up and running quickly has folks like Anthropic making large commitments to SpaceXaiX or whatever he'll wind up calling it. And then as public opinion continues to sour on ground based data centers he will step up as also the only viable option to get them built in space.

I think Anthropic's numbers will be pretty solid as they start to face the light of day. The question will be are they able to sustain that as the subsidy era and the tokenmaxxing eras both end. The former is good for them and the latter isn't. But I think the latter achieved what they were aiming for and companies are now hooked. OpenAI has a chance, and has been taking it, to gain some ground back on Anthropic as the subsidy life fades away and people actually have to start paying for what they use, but I do think Anthropic has a pretty comfortable lead right now even if Codex is making good strides.

The one thing that could and perhaps will hurt them is if Google continues to put effort behind the Gemma models as a way to compete with the Chinese open source models. Most enterprises will run to it as a way to save on token cost and that will take the wind out of the OpenAI/Anthropic sails.

274 views

I finished my build 2 yrs ago and outfitted my rig with 4 x 16gb Hynix RAM modules and I think I paid $230-something.

The same 4x16 today costs $800 after taxes.

The datacenter AI race all but wiped out PC RAM companies as Micron has already announced that they're focusing on datacenter RAM going forward.

230 views

How many millions of Blackwells are sitting in warehouses? Is there enough data centers to keep up with the demand to generate the income the AIs and Nvidia need to?

--
Freeman's got so much riz they need to call him Aura Parseghian

270 views

I would likely be typing this from my yacht.

AI Data Center Buildouts are going to be one of the biggest issues in the next few elections. Demand is through the roof, but NIMBY forces will be strong.

Maybe all those Blackwells that need a home is why they are starting to push for the distributed house-attached mini data center. Provide some of that excess energy from your home's solar panels and help someone create some AI slop no one at work will read. And maybe get a tax write-off to boot.

266 views

the thing you don't want in your backyard is extractive, unbridled natural resource depletion so that the billionaires can be trillionaires.

233 views

I wasn't trying to be flippant in a "that cell phone tower is ugly" sort of way.

It really is going to become an election touchpoint that all the usual worst politicians will twist to their end. It has tremendous potential to help local employment. It has even more potential to ruin towns.

I don't trust our current political landscape to have nuanced and open conversations on the best way we can achieve the economic benefits that can come with DCs balanced against the multiple negative consequences they also bring.

120 views

enriching themselves at the expense of the well-being of the other 99.9999999999999999999% of people.

102 views

to Memphis, which I think is great so they can discuss the social justice impact of the lightly regulated "mobile" power generators that have been parked in poor black neighborhoods and are powering data centers. I've been bombarding the youth minister with articles every day. I really want these kids to understand the big picture. Right now, they just see the "whiz-bang."

151 views

Assume the AI Data Centers required to power the levels of revenue required by OpenAI and Anthropic simply appeared on earth (or space) today. Would we be able to afford to operate them?

--
Freeman's got so much riz they need to call him Aura Parseghian

136 views

The fact that Google is planning to build a data center in my figurative backyard, where it will guzzle my community's drinking water, keeps me up at night. Personally, I don't want these things in anyone's backyard, but to Pat's point, plenty of people seem perfectly content to use AI, provided the environmental costs are unseen and borne by someone else's community. It's not unlike those videos of graduating students booing commencement speakers for celebrating AI: most of those students were happy to use AI to write their papers for them, but they're outraged now that the same tech they outsourced their work to might be bad for their career prospects.

Of course, all of this is part of a wider conversation about the real costs of AI and the Faust-like ways we're making our Faustian bargains with technology, but very few people seem willing to broach such matters these days.

166 views

It's not clear to me that the college students booing AI are private superusers, or if they are, that they are happy about it.

The idea that consumers are somehow implicated in this push for AI is just not accurate. I don't know anyone who is paying for AI or who has suggested that they would. We have been told there are strong business cases for its use to write code quickly or run predict consumer habits. Great. Let Pfizer game out how to manipulate public opinion while they also run data mining projects to find cures for cancer.

The private use case for AI remains incredibly weak if you can actually read and believe in your own moral worth. Even if you've done it, most people can transparently see the shittiness of turning in school essay without doing any work. If you tell that person that as a part of the bargain of being able to guiltily cut corners they have to accept being unemployed, enrich billioniares, and ruin their local and global environment, I would guess that most people would be more than willing to clarify their position.

This isn't a matter of personal moral confusion. When people can see what's happening they know what they want. It's a hard push by concentrated capital to force or convince everyone else this is good for them.

Also, this gets compared to the development of the internet. I don't recall the internet being massively unfavorable or having papal encyclicals written about it. Yes it is a technological revolution, but the terms are so different. Just one key difference is that the internet looked expand personal autonomy and association and AI appears designed to crush personal agency and diversity.

87 views

looked it up:

https://www.vatican.va/content/john-paul-ii/en/apost_letters/2005/documents/hf_jp-ii_ap...

14. The apostle Paul has a clear message for those engaged in communications (politicians, professional communicators, spectators), “Therefore, putting away falsehood, speak the truth, each one to his neighbor, for we are members one of another… No foul language should come out of your mouths, but only such as is good for needed edification, that it may impart grace to those who hear” (Eph 4: 25, 29).

To those working in communication, especially to believers involved in this important field of society, I extend the invitation which, from the beginning of my ministry as Pastor of the Universal Church, I have wished to express to the entire world “Do not be afraid!”

Do not be afraid of new technologies! These rank “among the marvelous things” – inter mirifica – which God has placed at our disposal to discover, to use and to make known the truth, also the truth about our dignity and about our destiny as his children, heirs of his eternal Kingdom.

86 views

Your average golf course uses far more water. Some of the early data center complaints on water usage trace back to a report that overstated usage by 10x due to a math error.

Electricity use, on the other hand... I'd be leery of a data center that's not showing up with its own generation capacity.

100 views

Assuming the center's daily water use ends up on the higher side--and I'm inclined to take the over--my city and its environs will need a new water source three decades earlier than originally projected. That fits my definition of "guzzling."

99 views

A fair concern, depending on the size of the water district supporting that.

Golf courses tend to run between 300k and 1M/day.

56 views

- No text -

38 views

powered by my little forum