What bugs me is how you hyperclouds and hysperscalers take a holier than thou position in ai coding simply because your tpu or titanium servers can't keep up with the demand of compute

Here’s a detailed section-by-section breakdown: HBM/DRAM Cartel (0:00) HBM is the #1 bottleneck right now. Only SK Hynix, Samsung & Micron control supply — basically a cartel. HBM3E is ramping slowly, HBM4 delayed. Demand from every big GPU cluster massively outstrips supply. Expect tightness for 2-3 years. DRAM slightly better but still constrained. Silicon Photonics & CPO (7:24) bubbleboi is bearish near-term. Indium Phosphide lasers (esp. from Lumentum) are severely supply-constrained. CPO complexity, yields, and qualification cycles mean copper/electrical interconnects stay dominant for years. Advanced Packaging – TSMC vs Intel (11:35) TSMC dominates and acts like a warlord — capacity oversubscribed. Intel has real potential in advanced packaging (EMIB/Foveros) and could become a meaningful challenger. bubbleboi has genuine conviction here. HBF & Memory Tech (13:49) Hybrid Bonding gives Sandisk/SK Hynix a short-term monopoly window. Ties into CXL for future memory pooling. Memory Accelerators (17:02) Solutions like TurboQuant that compress/optimize memory usage are practical near-term wins while waiting for more HBM. PCBs (22:35) Specialized high-frequency PCBs (e.g. Unitika) are an underrated choke point — tight capacity and long qualification. Power Delivery Shift (27:27) Big transition from 48V → 800V coming. Huge opportunity for SiC/GaN players (Wolfspeed, TI, etc.) repurposing EV lines. Other Back-end Bottlenecks (31:32) OSAT, fiber coupling, assembly & test — fragmented and full of hidden delays. 2-3 Year Memory Outlook (36:23) HBM stays tightest & priciest → DRAM improves relatively → NAND less critical. Alt Accelerators (44:10) Etched, Taalas, MatX etc. have real shot on inference side. Training still Nvidia’s moat. Cerebras post-IPO success is telling. Physics Walls (51:48) Voltage wall Cooling wall are real long-term limits — not fixable by just shrinking nodes. Rapid Fire Takes (56:30) • Intel: strongest conviction pick • Nvidia: training dominant, inference contestable • TSMC: still the industry tempo setter • Hyperclouds likely the future infrastructure model bubbleboi’s core message: AI progress is 90% brutal physical supply-chain reality, 10% software magic. Real alpha is in understanding the boring bottlenecks. Highly recommend watching the full thing if you’re in AI infra or investing in hardware/semiconductors. One of the best interviews on the topic this year. (Thanks @chrisbarber for hosting!)

I interviewed @bubbleboi about his ratings of AI supply chain bottlenecks. We talked about DRAM, advanced packaging, CPO, HBF, PCBs, power delivery, etc. 0:00 HBM, DRAM, the cartel 7:24 Silicon photonics, CPO, Lumentum lasers 11:35 Advanced packaging: TSMC vs Intel 13:49 HBF: Sandisk/SK monopoly window 17:02 Memory accelerators TurboQuant 22:35 PCBs Unitika 27:27 Power: transition from 48V to 800V 31:32 Outsourced assembly test, fiber coupling 36:23 HBM vs DRAM vs NAND in 2-3 years 42:03 Where will hardware founders come from? 44:10 Alt accelerators: Etched, Taalas, MatX 48:43 Emerging tech: CPO bearishness 49:23 Hyperclouds 49:53 HBF timeline CXL 51:48 Voltage cooling wall 56:30 Rapid fire: Intel, Nvidia, TSMC, Alphabet 1:05:25 ASML, Hynix, Lumentum, Wolfspeed 1:13:25 Building conviction in Intel 1:19:37 Pitching Intel to funds 1:22:39 X accounts, analysts why care about any of this

Even just using bare metal providers is sufficient to get a substantial, and I mean substantial, discount compared to hyperclouds. And if you use committed resource spend, you can even get further, like cents on dollars.

But they’ll all list on US markets, be supplied by US hyperclouds and chip makers, get funded by US VC, and many of them will move HQ to the US.

Problemem jest tempo innowacji i wyścig kto najszybciej będzie miec najlepszy model. Gdzieś mi mignął na x post że 1) proces projekt - akceptacja - produkcja asic dla hyperclouds trwa 2 lata, juz na starcie jesteś w sumie 1.5 generacji w plecy

by the way we have examples how to connect to your MySQL database hosted in all major hyperclouds developers.cloudflare.com/hy…

Special Oracle PR Operation pretend Oracle annual CAPEX is quarterly spend to make glorious Oracle belong with hyperclouds. Annual CAPEX number is still pathetically small in comparison with real clouds. accelerationeconomy.com/clou…

Shared in the Fission Discord: never thought about decentralised compute in this way, really amazing. Maybe its the example that's very attractive if not the theory. from youtube.com/watch?v=bo5WL5IQ… #decentralized #compute #hyperclouds

When $ORCL says they "are opening new and expanding existing cloud datacenters very, very rapidly", they are bonsai league compared to the hyperclouds. Clown, not cloud.

Hyperclouds on ##GalaxyS24Ultra. I will try if there's a clear sky to get a hyperl at night with 🌟 trail on. I tried to do it once on S22U, but I had no tripod, and I couldn't just stay 2 hours not moving with the phone pointed at night sky (go figure). I love this mode. I will play more on sunsets and whatnot. Maybe also play with astrophoto in ExpertRAW

All three hyperclouds see slight acceleration from previous quarter.

Looking forward to updating the x-axis after Oracle reports. Pulling away from IBM is one thing, competing with the hyperclouds another. And can always break out the cumulative spend chart.

Footsteps are good in Multiplayer but Warzone def subpar from the limited experience I have had… I have Hyperclouds to so it ain’t equipment my guy

How do the growth rates of $DLR and $EQIX stack up to the hyperclouds? Not so well (even as the data center companies are doing acquisitions while the clouds are an order of magnitude larger). /4

Managing game clouds are getting so complex that they’re already pushing the limits of traditional business cloud and data center management capabilities. Read the following @Spirent blog and see how #AI and #ML are the future of gaming hyperclouds. okt.to/QW937i

What’s weird is the hyperclouds basically exclusively want margin on compute — so they’re incentivized to make these things approachable — but it’s just not in their DNA whatsoever. Zilch. You can literally feel the org chart when you use AWS.

3 hyperclouds and 2 IEMS

The three companies with hyperclouds – Amazon, Google, and Microsoft – collectively spent over $124 billion on CAPEX in 2021, up 28% from 2020’s almost $97 billion.

⁦@hypecloudclub⁩ these will fly high then higher ⁦@1GoddRich⁩ ⁦@apocalypticapes⁩ taking over the skies and web3 #clouds #hyperclouds what!? 👀