4drant (@4drant) header
4drant (@4drant) profile picture

Tech Investor. everyone and everything here is a joke!

Mars
Joined December 2019
Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)Photo by 4drant (@4drant)
Pinned Tweet
The implications of KV cache offloading (NAND) and engram memory leading to a bull case on $PSTG (tail latency!). I cannot put it more succintly Thats right, Macron, packets need to be predictable, for sure!
1
1
1
22
12,233
200B from 1GW only off you spend 3GW in training Worth that training, perhaps is 15-20B/GW $ant Is it that we don’t need compute? Or that ANT is over earning
It's fascinating that the revenue density of serving inference compute to paying customers hasn't triggered more debate over the datacenter/compute capex market. Gerstner is effectively telling u that Ant disclosures show they are so revenue dense on inference that they can support $200bl in rev per gw at today's api/sub prices. The purpose to make the point that Ant economics are quite amazing at scale. But this argument can be flipped quite brutally in the other direction.. U could argue that at current API prices of the best frontier model you really only need 1gw of inference compute to serve all global frontier demand and then some. That then begs the question of what exactly is everyone else blowing zillions on to not be able to match this capability. I mean you could basically call wasted training massive shadow serving inference capacity. But why stop trying to catch ANT, if once you do you only need a gw to serve $200bl? But then doesn't that mean the training lead is existential and can't be seperated if fuck all compute can serve the planet once u ahead? Also, what happens at some level of capability saturation? Clearly if 1gw can serve 200bl in arr that number could collapse if the quality lead goes away. Either way think its kind of becoming obvious that compute assumptions might have some serious holes in them..
6
466
How long Colette survives at $nvda
57%3 months
4%12 months
14%She the boss
25%Results
28 votes • 1 hours
238
Excesses will be in DC (shell , energy, permitting)… not semis So maybe the delays by dems won’t matter
AI Overbuild hypothesis proven. Exhibit 100 To serve nearly 18.8B of revenue Anthropic needed only 500 MW of compute!!! Now imagine how much you can serve with just 10GW or 50GW. Per SemiAnalysis we were supposed to get 92GW by year end 2027, with 50GW being added in 2027 alone. They have revised it significantly downwards on Capex, but it is still very high.
1
2
440
How is he a bull?
Brad Gerstner at TomorrowX on Anthropic's Compute Efficiency: - Anthropic might only need 500 MW of inference next year to serve their customers
4
387
All supply side analysis is BS What you need is a good productised always-on agent and we get the Claude code like demand inflection and we are massively short compute again This will happen in the next 6-12 months, I don’t know when
1
2
1
13
2,537
What do you do when the most bullish analyst cuts numbers? Just do the same
Semis should be fine. $NVDA and its partners are in trouble
2
390
Or is SA cut making the bottom?
SA’s latest downward revision is a bit of a shock… Is the AI trade over?
6
59
5,846
Is it the same guy who wants dram out of your pc at home? And calls you irresponsible Yes I trust him
Sam Altman told Samsung executives AI demand has been "double counted" amid neocloud bubble concerns. He said only OpenAI and Anthropic drive major demand growth, warning of supply shocks and turbulence, but expects AI demand to surge long term. Source: fnnews h/t @jukan05
1
1
469
Dan Durn for $nvda ceo Who is with me?
Who do we want to be CEO of $NVDA?
279
Do they realise they are posting this after the official number is out … and this shows how useless their tracker is?
NEW Ramp AI Index: In October, AI spend rose 2%. OpenAI and Anthropic growth fell to low single-digits. Engineering teams drove 72% of spend, followed by customer service and leadership. Given this relatively modest growth in spend, and new competitive dynamics that may stick around, the markt should look to new fronts in the AI trade.
4
53
3,937
Citrini bears are bearing around Ofcourse you can’t believe and it’s a load of crap but it’s also not as bad as you might have made it appear
I will be really surprised if they can rent out Ampere/Hopper at such an insane premium in 2027/28, even when Rubin and Rubin Ultra will be in full production The only scenario in which this happens is if DC construction slows down a lot due to regulatory reasons
1
411
Great guidance, now they can beat it They are learning $ant
Apparently, Anthropic gave conservative Q4 guidance because engineers tend to take time off in December. Seems like a sensible approach.
6
1,057
What do you expect when we have Blackwell and Vera?
Happy Monday! The H100 forward curve just flipped into contango. For the first time since Ornn started tracking it, the 6-month term is pricing over the 1-month. Do we think this is a trend or a blip?
4
433
Ofcourse everyone knew about it already Number of pictures he clicked
黄仁勋和情人Alice Huang更多照片首次曝光
3
539
At this point, it makes 100% sense for $avgo to bid $1000 cash offer on $lite even if it may not cross anti trust It makes the deal difficult for $mrvl to do You think it happens?
UBS: "Another buyer could come over the top for $LITE"
6
1
1
39
8,188
It becomes very unreliable, several times server times out or no completions
Operation Cheepseek Phase 4: $100 of usage for DeepSeek V4.1 Flash
2
310
$goog Find the mistakes
311
No $AAPL shareholder seems to worry
I don’t think most ppl will carry a phone around in 10 yrs.
1
1
719
Pulse check
7%Selling
53%Buying
40%Frozen
30 votes • Final results
249
Even this gap has filled What Are You waiting for? $wdc
$WDC If this setup doesn't hold, watch out below.
2
718