Sabitlenmiş Tweet
If you can sleep at night, you aren’t trading big enough. #trading #quantitativetrading
English
Tail Risk
3.6K posts
Tail Risk retweetledi
The AI investment cycle is only accelerating:
Global data center CapEx driven by AI is projected to reach $5.2 trillion by 2030, according to McKinsey.
IT equipment would represent ~$3.3 trillion of that total, followed by data center infrastructure at ~$1.6 trillion and power generation at ~$300 billion.
This assumes 125 incremental gigawatts of new AI data center capacity added between 2025 and 2030, requiring as much electricity as ~125 nuclear reactors to power.
In an accelerated demand scenario, total CapEx could rise to $7.9 trillion, with 205 incremental gigawatts of capacity added.
A constrained scenario would require $3.7 trillion, with 78 incremental gigawatts added.
The investment is expected to be driven by mass adoption of generative AI, enterprise integration across industries, competition between mega-cap tech and other firms, and governments investing heavily in AI infrastructure.
The AI buildout is set to reach unprecedented scale.
English
Tail Risk retweetledi
i'll bet you 6 quintillion dollars that quant trading is not solved
Minara AI@minara
Quant trading is solved. Introducing: Minara Strategy Studio. Design strategies, backtest, compare PnL, avoid brutal drawdowns, optimize parameters, go live, all in plain words. Join the closed beta: strategy.minara.ai 🧵
English
Tail Risk retweetledi
$AMZN CEO: “If you are building a big inference business and want decent margings, not having your custom silicons is a structural disadvantage.”
Customers doesn’t just want better performance, they want better price/performance.
Currently, we can’t deploy or use AI as expansively as we want because inference is still expensive. They run mostly on expensive $NVDA chips.
The whole industry will eventually move away from $NVDA to better price/performance alternatives. This will accelerate when gains from training plateaus and most of the workloads shift to inference.
$AMZN is well positioned to offer the best price performance as they are using their custom CPUs and XPUs at scale and the largest AI labs, Anthropic and OpenAI, are already their customers.
Cloud growth is already expected to reach over 30% this year, proving that their strategy works.
The market is not bullish enough on $AMZN.
English
Tail Risk retweetledi
Every conversation I have with @dylan522p, I'm really just trying to understand the supply and demand of tokens.
This is a unique episode in that it's entirely dedicated to talking about both sides of that equation.
We discuss:
- The infinite demand for the newest models
- @SemiAnalysis_ going from $10K on AI spend to $7M
- Mythos and Anthropic's compute problem
- Why TSMC spending $100B on CapEx could cause a shortage
- Robotics as next demand wave
- Why memory prices will double again
This is my second conversation with Dylan and find myself needing to speak with him more and more often to make sense of it all.
Enjoy!
Timestamps:
0:00 Intro
1:00 Surging AI Spend
10:27 Token Demand
16:21 When Ideas Are Cheap and Execution is Easy
20:46 Model Hoarding
22:34 Robotics
27:03 The Compute Bottleneck
30:26 The AI Permanent Underclass
31:39 Supply Chain Reality
37:47 CPUs
42:54 Predictions: Public Backlash
English
else helping musk build terafab seems like digging their own graves no? terafab can pretty much do everything $intc does and more.....
Tail Risk@_tailrisk
wouldn't it make sense for TSLA/SPACEX/XAI to just acquire intel?
English
wouldn't it make sense for TSLA/SPACEX/XAI to just acquire intel?
Elie 📽🎬(Melanated Ventures)@ElieJlouis
$INTC Super bullish. This was from $TSLA Q1 earnings call about Intel involvement in the #terafab #14a 🔥
English
Tail Risk retweetledi
Hide your Defi Protocols
Mythos is coming to town
zerohedge@zerohedge
*ANTHROPIC’S MYTHOS MODEL ACCESSED BY UNAUTHORIZED USERS
English
Tail Risk retweetledi
Wow, KelpDAO comes out and says:
> 2 of LayerZero’s RPCs were hacked
> it was LayerZero internal compromise that led to the exploit
> they took fast action to prevent another $75m vulnerability
> the 1/1 DVN was the suggested setup from LayerZero & even after they asked further about it during the transition to L2s, it was kept the same
> blames LZ for the setup
My goodness. Absolutely no one taking any responsibility and no real detail on the loss socialization for Aave users still.
I think we are all underestimating how long the WETH & stablecoin pools may be frozen.
Kelp@KelpDAO
English
@MikeFritzell Why do you say it’s a bull market? Just from the tick up in prices? Just curious to get your thought.
English
The Hong Kong property bull market has begun
English
Tail Risk retweetledi
I'm dropping a thread of all the protocols that had to freeze their interop because of LayerZero being compromised.
Let's go:
English
Tail Risk retweetledi
Here's how to triage:
1. Go to admin.google.com
2. Security → Access and data control → API controls → App access control → Manage Third-Party App Access
3. Search for client ID:
110671459871-30f1spbu0hptbs60cb4vsmv79i7bbvqj
if found → revoke / block
Vercel@vercel
Our investigation has revealed that the incident originated from a third-party AI tool with hundreds of users whose Google Workspace OAuth app was compromised. We recommend that Google Workspace Administrators check for usage of this app immediately. #indicators-of-compromise-iocs" target="_blank" rel="nofollow noopener">vercel.com/kb/bulletin/ve…
English
Tail Risk retweetledi
The chip is dead
Long live the factory
NVIDIA no longer ships a GPU. It ships a factory that gets better after it leaves the loading dock
2.7x more productive in 6 months
On the same silicon
Without changing the hardware
Wrote up the full thesis $NVDA
bepresearch.substack.com/p/the-chip-is-…
English
Literally 6 days ago. This likely has nothing to do with mythos... but i mean defi is unusable. All downside, sitting ducks.
Tail Risk@_tailrisk
is anyone worried about mythos opening the floodgate for enabling hacks across defi? I'm concerned and I want to pull everything. If you aren't worried, find a way to short defi?
English
Tail Risk retweetledi
Tail Risk retweetledi
How does memory determine inference speed?
What are the types of memory?
Why did Nvidia buy SRAM based Groq?
What do AI processors do?
Processors do three things.
1. they calculate results (Core).
2. they store the results (Memory).
3. move results to where they are needed (I/O communication).
What determines Inference Speed for a Processor?
Understanding speed is just understanding the bottleneck in each architecture. It’s a sort of engineering “Where’s Waldo.”
The calculation in AI inference is trivial. Its mostly multiplication and addition. Processors in general are very good at multiplication and addition.
Calculation is not the bottleneck.
In fact, graphic processing units, GPUs, are so fast at calculation that they are often sitting idle waiting for the data required for the next calculation. This is an important hint.
The reason GPUs can’t do fast inference is that the GPU cores can’t get data fast enough from memory.
This isn’t the core’s problem. It’s the memory’s problem.
What are the types of memory?
There are two types of memory.
Memory that can store a lot, but is slow at moving data.
And memory that is fast, but can’t store very much data.
The former is called DRAM (or HBM) and the latter is SRAM.
Graphics Processing Units use DRAM.
Why?
Because graphics was the perfect use case for DRAM.
In graphics, data is moved from memory to compute occasionally, and the calculation phase takes a long time.
So it didn’t really matter if the movement of data was slow. The calculations took so long that the time it took to move data gets lost in the wash.
But AI inference has the exact opposite characteristic.
The compute time is small.
While data moves constantly from memory to compute.
In AI inference, the time it takes to move data from memory to compute is the lion’s share of the total time.
And this is exactly where DRAM is slow.
Lets see DRAM/HBM in action:
To generate one word in an answer, all of the weight data need to move from memory to compute.
And the weights are very large.
Like hundreds of HD movies worth of data.
Here is the kicker. The process is serial.
So for the next word, all the weight data needs to again be moved from memory to compute. And again and again. For every single word in the answer, hundreds of HD movies worth of data need to move from memory to compute. Because DRAM/HBM is slow, moving data is time consuming.
While the DRAM/HBM is moving data, the GPU is waiting.
It is sitting idle. Pulling power. Doing no work.
This is why GPUs can’t do fast inference.
SRAM on the other hand moves data extremely quickly.
So it is good exactly where DRAM/HBM is bad. It can deliver data extremely quickly to compute.
This is why all high speed inference solutions use SRAM. But… and there is always a but.
But SRAM can’t store much.
And the weight data is very large and needs to be stored near the processing core.
There are two options...
What are the 2 options to solve DRAM/HBM’s weakness?
If you want to use SRAM you really have two choices.
1. You can use hundreds or thousands of little chips. Each with a little bit of SRAM and a little bit of compute. This was Groq’s strategy.
2. Or you can use very big chips. This was Cerebras strategy.
If you build an SRAM solution with little chips you need to break up the weight data, and put it in the SRAM of thousands of little chips.
You divide and conquer.
But then you need to link them together so they can communicate with each other.
And this is where IO/communication comes in and rears its head, because chip-to-chip communication is slow.
Thousands of little chips communicating over cables and through switches is slower and more power hungry than if all that traffic stayed on a big chip, or even several big chips. Communication between chips is slow, and communication on-chip is fast.
So some of the gain won by using SRAM is lost as the hundreds or thousands of little chips need to communicate with each other hop by hop by hop.
This is why wafer scale was the holy grail.
It provided the benefits of SRAM while mitigating SRAMs weaknesses.
By building a chip the size of a dinner plate, a chip that is 58x larger than the largest GPU, the designers could stuff it to the gills with SRAM.
One can’t make SRAM store more data per square millimeter, but one can provide more square millimeters by building a bigger chip.
And by building a bigger chip one avoids nearly all of the cost (in time and power) of communicating across chips.
This is what we achieved at @cerebras. And why we are faster than Groq.
And why we chose to build a wafer scale processor. And while we still sometimes need to communicate across chips, it is only across 4 or 8 or 12 chips rather than hundreds or thousands.
The result is that Cerebras is the fastest AI processor in production.
English