Jacob Kilpatrick

66 posts

Jacob Kilpatrick banner
Jacob Kilpatrick

Jacob Kilpatrick

@_killyp_

Katılım Haziran 2011
208 Takip Edilen67 Takipçiler
Jacob Kilpatrick
Jacob Kilpatrick@_killyp_·
@jrysana @bubbleboi But it is the way of the future in AI compute. Dedicated chips, not GPUs. And they have a massive head start being on their 8th generation.
English
1
0
0
34
bubble boi
bubble boi@bubbleboi·
People underestimate xAI at their own peril. They’re the real wild card in the AI race, and the bigger picture reveals an actual strategy to not just leapfrog the competition but to become the only winner in AI. It’s eerily similar to the playbook John D. Rockefeller ran with Standard Oil. If you look at the funding picture today OpenAI just raised $122B (that they’re lighting on fire) and Anthropic has raised ~$72B in its entire lifetime. xAI in comparison has raised meaningfully less than either. But xAI somehow already has the most compute online right now, and the capital structure is about to flip that’s going to turn them into one of the largest free cash flow machines in history. SpaceX filed its S-1 last week targeting a ~$75B raise at a 1.72 trillion valuation and although most of that is going to go to Starlink and Starship, the detail many are missing is that xAI now sits inside a public-company balance sheet with access to huge institutional debt markets that pure AI labs don’t have. OpenAI and Anthropic have to keep selling equity to fund compute while xAI can issue investment-grade paper against Starlink’s $4.4B in operating income, giving it a dramatically lower effective cost of capital which is critical in this gold rush to expand compute infrastructure and capture market share. If you then look closer at the Colossus deal with Anthropic xAI gets ~$40B to rent out their old infrastructure which allows them to monetize a depreciating asset at a higher rate than they were getting from internal use, as well as control a competitor’s inference economics, and recycles the cash into Colossus 2. This is a Rockefeller move no doubt. What many don’t understand is that Standard Oil didn’t just refine oil. It owned the rail rebates which in turn made their competitors pay them every time they shipped a barrel. Rockefeller’s rivals literally funded the monopoly’s expansion and similar to today with a severely supply constrained compute market anyone with cheap power right now can pull such a move (Colossus taping into the Natural Gas pipeline was another genius move for many other reasons). This matters a lot more than people think right now because we are still in the blitzkrieg part of the AI build out but as we shift from training to inference, compute capacity will directly equal revenue capacity. The pre-training benchmark game the AI labs are playing right now is hitting marginal returns. The customer-serving game is just starting, and whoever has the most power online and the best economics wins it. When you factor in the SpaceX IPO, public debt access, $40B of incoming Anthropic revenue (and many more deals to be announced soon I believe) as well as the first operational gigawatt cluster with Colossus 2 we are going to see a sea change in the industry. Power, uptime, and compute access will be all that matters. If I am right I expect consolidation with the big AI labs and AI services. The recent mergers & buyouts with SpaceX are just the start.
bubble boi tweet media
English
50
43
656
92.4K
bubble boi
bubble boi@bubbleboi·
I’m turning increasingly bearish now. Have a feeling around October we will see something really bad.
English
121
18
1.1K
158.9K
bubble boi
bubble boi@bubbleboi·
You think the AI bottleneck is HBM. You think it’s CoWoS. You think it’s GB200 cables or 800G optics or the Arizona power grid. You are looking at the wrong layer of the stack…. The REAL bottleneck is epoxy resin paste. Specifically, Liquid Compression Molding compound EME-G, a goopy, beige, photosensitive thermosetting resin that gets squeegeed onto HBM stacks before the mold press comes down and cures it. Without this paste, the silicon dies in an HBM stack delaminate, the TSVs crack, and your $40,000 GPU becomes an expensive paperweight. Sumitomo Bakelite (4203.T) makes roughly 90% of the world’s supply. The other 10% is split between Nagase ChemteX and Hitachi Chemical, both of whom buy precursor chemicals from Sumitomo Bakelite. The moat is vertical. The resin formulation is a trade secret developed over 38 years of iteration. It contains a specific ratio of silica filler to bisphenol-F epoxy with a coefficient of thermal expansion tuned to within 0.3 ppm/°C of silicon. Get the ratio wrong by 2% and the HBM stack warps during reflow. Samsung tried to qualify a domestic Korean alternative in 2022. They failed. SK hynix tried in 2023. They failed. Micron didn’t even try. Sumitomo Bakelite ships approximately $180M of EME-G annually at a gross margin of 74%. Each HBM stack consumes roughly $6 of resin. As Nvidia’s roadmap points to more and more HBM stacks the math is clear. Feynman GPU package has 8 HBM stacks. That’s $48 of Sumitomo Bakelite content in a $70,000 GPU. They are 0.07% of the BOM and 100% of the bottleneck. This is the most asymmetric pricing power in the entire AI supply chain and they are charging like it’s commodity epoxy because the company is run by Japanese chemical engineers who think 8% annual price increases are aggressive. In Q3 of 2026 the Nvidia Rubin Ultra ramp is scheduled and will trigger an EME-G shortage. Sumitomo Bakelite will have to raises prices 35%. The stock will get re-rated from “specialty chemicals” to “AI infrastructure.” Multiple expansion from 14× P/E to 38× P/E. It’s a three-bagger in 18 months and the stock is up >3% YTD.
bubble boi tweet mediabubble boi tweet media
English
75
90
1.8K
668.3K
Josh Woodward
Josh Woodward@joshwoodward·
In @GeminiApp this is what you're working with now: Fast = 3 Flash Thinking = 3 Flash (with thinking) Pro = 3 Pro (with thinking) Our best line-up yet!
Josh Woodward tweet media
English
172
242
3.3K
324.7K
Jacob Kilpatrick
Jacob Kilpatrick@_killyp_·
@qauff @JackRhysider +1 for fish. It isn't as customizable as Zsh but it comes with so many awesome features out of the box it has become my go to.
English
0
1
9
0
Fifth Quarter 🍢FSU 🍢
Fifth Quarter 🍢FSU 🍢@fifthquarterfsu·
Hey guys, so there's a big outage on Discord - not just us, a bunch of other folks are down. They know about it and are taking care of it.
English
4
1
18
0
Jacob Kilpatrick
Jacob Kilpatrick@_killyp_·
@MalwareJake @Steph3nSims I used IDA over Radare2 because IDA was MUCH better. Now with Ghidra being publicly available, comparable or better than IDA and is free, I almost exclusively use Ghidra.
English
0
0
1
0
Jake Williams
Jake Williams@MalwareJake·
For people who complain about the price of IDA, do you use Microsoft Office or Libre Office? Microsoft Office you say? Why, when Libre Office is free? Exactly... @Steph3nSims #PenTestHackFest
Jake Williams tweet media
English
4
3
20
0
Maddie Stone
Maddie Stone@maddiestone·
What are you favorite keynotes from conferences? Even better if you can include a link to a recording! What made it so great? Help me learn, please!
GIF
English
5
7
29
0
Jacob Kilpatrick retweetledi
b33f | 🇺🇦✊
b33f | 🇺🇦✊@FuzzySec·
I spent some time studying Donut (github.com/TheWover/donut) recently, great work by @TheRealWover and odzhan <3. I wrote a small loader for testing, UrbanBishop, which is doing some interesting tricks. Details on GitHub => #urbanbishop" target="_blank" rel="nofollow noopener">github.com/FuzzySecurity/…
b33f | 🇺🇦✊ tweet mediab33f | 🇺🇦✊ tweet media
English
2
95
206
0
Jacob Kilpatrick retweetledi
Jake Williams
Jake Williams@MalwareJake·
Attacker buys a use-after-free zero day exploit to gain initial access, but then can't get psexec working for lateral movement and sets off alarms everywhere...
English
3
5
37
0