[BluSyn]

You seem to be assuming that the full cost of the cluster is recouped by Grok 3. The real value will be in grok 5, 6, etc…

xAI also announced a few days ago they are starting an internal video game studio. How long before AI companies take over Hollywood and Disney? The value available to be captured is massive.

The cluster they’ve built is impressive compared to the competition, and grok 3 barely scratches what it’s capable of.

[Tycho]

Yes. Why do get these replies on HN that seem to only consider the most shallow, surface details? It could well be that xAI wins the AI race by betting on hardware first and foremost - new ideas are quickly copied by everyone, but a compute edge is hard to match.

[HarHarVeryFunny]

The compute edge belongs to those like Google (TPU) and Amazon/Anthropic (Trainium) building their own accelerators and not paying NVIDIAs 1000% cost markups. Microsoft just announced experimenting with Cerebras wafer scale chips for LLM inference which are also a cost savings.

Microsoft is in process of building optical links between existing datacenters to create meta-clusters, and I'd expect that others like Amazon and Meta may be doing the same.

Of course for Musk this is an irrational ego-driven pursuit, so he can throw as much money at it as he has available, but trying to sell AI when you're paying 10x the competition for FLOPs seems problematic, even you you are capable of building a competitive product.

[Tycho]

Timing matters. A long term strategy for superior hardware might bear fruit too late.

[HarHarVeryFunny]

I'm not sure about that - I expect AI is going to become a commodity market, so it doesn't matter how late you are if you've got a cheaper price.

In terms of who's got a lead on cheap (non-NVIDIA) hardware, I guess you have to give it to Google who are on their 6th generation TPU.

[vardump]

I wonder how Tesla's training computer Dojo is doing. Although I guess there's a reason for buying so much Nvidia hardware...

[theckel]

Curious where you saw the Microsoft/Cerebras experimentation noted online? That's very interesting.

[HarHarVeryFunny]

It was mentioned in Anthropic Jack Clark's "Import AI" newsletter.

https://jack-clark.net/2025/02/17/import-ai-400-distillation...

[gmerc]

DeepSeek just showed the compute edge is not that hard to match. They could have chosen to keep the gains proprietary but probably made good money playing the market instead, quants as they are.

https://centreforaileadership.org/resources/deepseeks_narrat...

If you’re using your compute capacity at 1.25% efficiency, you are not going to win because your iteration time is just going to be too long to stay competitive.

[scarmig]

Software and algorithmic improvements diffuse faster than hardware, even with attempts to keep them secret. Maybe a company doubles the efficiency, but in 3 months, it's leaked and everyone is using it. And then the compute edge becomes that much more durable.

[mirekrusin]

Optimisation efforts don’t negate investment in capacity but multiply output.

[Tycho]

Sorry, you missed the point - DeepSeek tried some new software ideas, they did not manage to secure the same computation capacity.

[gmerc]

They achieved the same results for 1.25% of the computation cost... If they actually had that computation capacity, it would be game over with the AGI race by the same logic.

[acchow]

> but a compute edge is hard to match.

xAI bought hardware off the open market. Their compute edge could dissappear in a month if Google or Amazon wanted to raise their compute by a whole xAI

[Tycho]

Not if there’s a hardware shortage.

[acchow]

Ok, 2 months.

Remember, the new B200 have 2.2x the performance of xAI’s current H100 “hardware edge”. So it only takes an order half the size.

Or you could order the old H100 instead and avoid the B200 shortage.

[kmac_]

The point is that it is inefficient. Others achieved similar results much cheaper, meaning they can go much further. Compute is important, but model architecture and compute methods still outweigh it.

[HarHarVeryFunny]

How quickly will Grok 4/5/6 be released? Of course you can choose to keep running older GPUs for years, but if you want bleeding edge performance then you need to upgrade, so I'm not sure how many model generations the cost can really be spread over.

Also, what isn't clear is how RL-based reasoning model training compute requirements compares to earlier models. OpenAI have announced that GPT 4.5 will be their last non-reasoning model, so it seems we're definitely at a transition point now.

[gmerc]

At current efficiency? Not nearly as fast as DeepSeek 4 ;)

[gmerc]

None of which explains this massive waste of money for zero gain.

[Larrikin]

It's not going to be from this unless it's forced upon us by the federal government. All the other companies are ahead and aren't just going to stop.

[doctorpangloss]

> xAI also announced a few days ago they are starting an internal video game studio.

Ha ha. I'm sure their play to claim airdrop idle game will be groundbreaking.