241
The Time Bomb Went Off: AI's All-You-Can-Eat Era Just Ended in Real Time (www.thestateofbrand.com)
submitted 5 months ago by to c/fuck_ai@lemmy.world
55 comments fedilink hide all child comments
 

Excerpt:

The IPO Math Forces the Issue

Both OpenAI and Anthropic are on IPO timelines for the second half of 2026. OpenAI completed the largest private funding round in history in April, $122 billion at an $852 billion post-money valuation. Anthropic has reportedly surpassed $30 billion in annualized revenue. Massive numbers, both of them. Also both attached to companies that are still burning cash at extraordinary rates.

Public markets will not tolerate the gap between subscription revenue and compute cost that has defined the past three years. The moment either company files, analysts will demand unit economics that show a path to margin. Usage-based billing is the fastest way to demonstrate that path.

None of this contradicts the repricing thesis. The pricing war is the last land grab before the gate closes. Both companies are spending aggressively now to lock in users whose switching costs will make them sticky when prices rise. OpenAI offers two months free. Anthropic offers 50% more capacity. Both expire in July. What comes after July is the real pricing.

you are viewing a single comment's thread
view the rest of the comments
[–] 13 points 5 months ago* (last edited 5 months ago) (2 children)

contract with an SLA

Plenty of hosters provide that. Cerebras, for example, fabs their own ASICs (seperate from Nvidia), builds them into servers, hosts a number of open-weights models themselves in friendly jurisdictions, and offers SLAs for enterprise clients; it doesn't get more "guaranteed" than that in AI Land, but there are tons of hosts to choose from.

https://www.cerebras.ai/build-with-us

The major sticking point is that the best open weights models are Chinese. This doesn't actually matter from a security standpoint anymore than buying a Chinese tire does; they're dumb weights anyone can finetune, host and run on whatever software/hardware stack one wants... But try explaining that technical distinction of "using Chinese AI" to executives responsible for entire corporations.

There are even attempts to "launder" Chinese models to make them palatable for western enterprise use. For example:

https://huggingface.co/microsoft/MAI-DS-R1

https://huggingface.co/unsloth/r1-1776

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 5 months ago (1 child)

    Plenty of hosters provide that. Cerebras, for example, fabs their own ASICs (seperate from Nvidia), builds them into servers, hosts a number of open-weights models themselves in friendly jurisdictions, and offers SLAs for enterprise clients; it doesn’t get more “guaranteed” than that in AI Land, but there are tons of hosts to choose from.

    This makes sense for first-party hardware businesses like Cerebras that are renting or selling their platform to developer businesses (second party) for the purpose of creating AI-based software tools which they will then sell as services to other businesses (third party), and I can see that guarantees could be written in a contract for the first-to-second-party relationship.

    What I don't see is that any such guarantees can be effectively written or enforced in a second-to-third-party contract, where an AI SaaS company is selling their software service to companies that don't do their own development, and expect that the service they have contracted will produce reliable results.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 2 points 5 months ago*

    Actually, what Cerebra’s does is no different than any generic host. They provide API access to LLM weights, though most providers will do it with some standard open source serving software like VLLM or SGLang.

    And they all use the same open weights LLMs. They arent the software developer.

    Cerebras doesn’t train their own model. And I think this is fine for service guarantees as long as the weights do not change, hence will provide the exact same deterministic results at zero temperature (and generally perform the same when used as a service).

    My experience is that a lot of “enterprise” LLM stuff is used in bulk, for results that can be “good enough” with a reasonable error rate. Like (for example) extracting info from literally millions of documents. Or as RAG/querying their own internal documentation.

  • permalink
  • fedilink
  • source
  • parent