Find it very interesting based on current narrative everyone is very confident we are already overbuilt. But the indications from those like ANT/Jassy etc is they cannot serve the demand and they are looking to build way more and Dario is even trying to be conservative and still has that much compute for the future. Seems very hard to square these when its all but training compute will be reallocated to inference/improvements in token throughput/algo efficiency/etc.
While I find the points you are making very thought provoking. How do you reconcile Dario who has been the most conservative on compute build out due to the potential demand mismatch having the size of commitments for compute through 2029 on the books? Clearly they are compute constrained.
Further cyber is clearly worried hence the PANW deal so we will definitely get an overbuild but not sure it is this wave instead of every Software co realizing they need to train their own domain specific model and need compute so the ecosystem overbuilds but I’m not sure all these firms have realized when you still see them announcing partnerships with ANT and OAI
Sure but with Anthropic supposedly solving distillation that may not be the case for long and it would likely make sense for the US gov to push them to help other frontier models solve that to entrench US lead.
But without it you want agentic cyber to have stronger models able to harden everything and defend against an inferior model that is always chomping on the heels. If you halt China isn’t just going to.
Argument for not releasing newer models but they really need to just harden everything ideally and then hardware is harder
Curious as I have read your article a few weeks ago.
If I remember correctly you had the reason as not only improved chips producing more tokens per watt but also the shift of compute from training to inference but it seemed to be a fairly aggressive shift modelled any reason for that speed or you just think no new compute really allocated to training and all new GW are inference?
I think a lot of what I have seen for the glut bear case requires a large pivot from training to inference compute. Maybe you see other cases like algo efficiency but that is the big one with the improvement in token output.
However if we are yelling about coding not having a high enough TAM makes it seem unlikely that we don’t see a slower transition from training to inference compute as they tunnel new domains. Could be very wrong.
Why do you believe they are over earning?
If frontier economics suck and we end up in a situation where financing for the neos dries then really most demand settles at the CSPs and they have the installed compute base even with margin compression they would absorb volume on increased supply no allowing a similar level of earnings and growth to potentially persist
Not demand side numbers but seems that CIOs have just been letting it happen but I am sure when you count Shadow IT (if Enterprises can get that those use cases compliance approved) there is pretty strong ROI and demand to roll out especially with better enterprise specific harnesses.
@ApresCapital@DratchCap I believe @BenBajarin mentioned on an episode of The Circuit. Most CIOs arent even tracking but those that are are mostly seeing some it seems. https://t.co/gacS1RAYza
I dont understand the path to lower capex.
Data center moratoriums - but jurisdictions compete for capital and not all will ban. If there is enough that do - other countries will get that investment (AB, Canada for example).
The hypers are all still dealing with the red queen of game theory.
Agree maybe I’m an idiot but I don’t see how cyber isn’t monstrous if you need frontier to protect in an always on manner against the non frontier open models that trail. 24/7 agentic defence will be huge as basically they have created an issue that will tax the system via higher cyber budgets
@rennyzucker@benjamin_horne Did you use codex cloud at all? Would love to port everything I’m doing but Claude code seems like better harness for non local work instead of super small prs
@_grojo@PythiaR Sure but I mean 90%+ of wealth groups young people aren’t doing anything. Avg advisor is a mid 60’s male who is tech illiterate. Most young advisors are trying to build a book an haven’t touched codex or Claude code.
Don’t think X is an accurate representation of the base
@6d617474@shortbus_ace@evrgn11112231 Worked at 2 banks it is amazing how bad they are at it and their hate of cost centers. I was legit asked by senior mngmnt at one (both >$100Bn mcap) how other was doing something I explained and they just cant wrap their mind around it even though it was good for 2 sep. divs
@shortbus_ace@evrgn11112231 Stretch to say TM enables multi-banking. TM is a proprietary value-prop product that used neutral rails. Banks feed standard balance data (which they give everyone, incl QB) but never share prop data/distribute each other's portals. Coop at the rail layer, not the product.