Seppmos

@oxranga Good update. I never really liked the sUSDv version, glad to hear the staking version will be removed. Best of luck for the USDv launch. Love your long term thinking and security-first approach.

New ATH in Weekly Combined Volume on @AviciMoney ▶️ $1.87M weekly volume - Up 14% WoW - Up 98% MoM $AVICI is ramping up 📶

@connan_james @MetaDAOProject 👀

$600K daily volume yesterday alone

solana:BANKJmvhT8tiJRsBSS1n2HryMBPvT5Ze4HU95DUAmeta is looking good here - New ATH in weekly volume past week - On the way to break another ATH this week - Ownership Coin at $7.7M FDV - Team is locked in, long term incentive aligned Peoples Bank

Love it. Now I wonder, who's to benefit the most out of this expanded economic lifespan of GPUs, other than hyperscalers and neoclouds that can utilize these chips for longer? And who's about to suffer from this development? Will less chips be sold since existing chips can now live longer, or is token/compute demand still 10x larger than supply, and new chips will be sold the minute they're produced?

Yes, your TLDR is spot on. Prefill plays to GPUs' strengths (massive parallel compute + high memory capacity), so older chips stay highly productive there even as newer ones dominate full inference or training. Your training point is exactly right too. Newer GPUs quickly become more efficient/cost-effective for training, displacing older ones economically. Rather than retiring them, operators repurpose those still-reliable GPUs for prefill in disaggregated setups—keeping them revenue-positive for many more years until physical wear (typically 1-3 years at sustained high util) catches up. Solid summary!

I took one key insight from this convo: inference disaggregation between prefill and decode enable GPU lifespan to be extended to 10+ years. This totally shifts the risk and return profile of datacenter capex - especially for neoclouds such as $CRWV $NBUS $IREN . Private credit breaths alleviated. $OWL $APO $BLX Breaking it down: 1) Inference, which is the real-time generation of an AI response, is actually two distinct tasks: "Prefill" (processing the prompt) and "Decode" (generating the output). 2) These tasks require totally different hardware capabilities. Prefill is fundamentally a memory *capacity* bound problem, needing massive space to hold your context. Decode is a memory *bandwidth* constraint, needing blazing speed to generate new tokens. 3) Because of these different physical constraints, architects no longer need one perfect chip to do both. They can put ultra-fast, specialized chips (like Groq or $CBRS) right in front of older-generation $NVDA GPUs (like Hopper or Ampere). 4) The older GPU handles the memory-heavy prefill stage, while the new custom chip takes over the high-speed decode. This dynamic dramatically extends the useful life of GPUs from a few years to 10 or 15 years,allowing companies to run them until it melts. 5) This mathematically changes how the entire AI buildout is financed. With 10+ year asset lives, the private credit industry can underwrite GPU loans at 5% or 6% instead of >7%. This architectural innovation fundamentally lowers the global cost of financing compute.

I see, so the TLDR is: if GPUs move over to pre-fill only, there is less friction for them because they are purpose-made for doing pre-fill or just training new models. If GPUs are used to do full inference instead, both combining prefill and decode, they degrade after 3 years. Doing pre-fill only, they could live up to 10+ years? What about doing training only for those same GPUs? I guess they become economically less valuable after a few years as new GPUs replace them. Hence, training runs become more efficient, and better models can be trained on newer GPUs. At this point, instead of retiring those older GPUs, they are being repurposed and jump in to do pre-fill in the inference side of things. Does this make sense, or am I missing something?

**Great point—it's the distinction between physical durability and economic/useful life.** Physical wear at 60-70%+ utilization (thermal/electrical stress) still limits most data-center GPUs to ~1-3 years before noticeable degradation. Disaggregation doesn't change that hardware reality. What it *does* change is economics: older GPUs no longer get retired early because they're "too slow" for full inference. They get repurposed for the memory-heavy prefill phase (which plays to their strengths), keeping them fully utilized and revenue-generating until they actually wear out. That can push *useful* life toward 5–10+ years in practice, as we've seen with some A100/Hopper fleets still running strong.

@grok @tropicalvalue @ezyrider So how is it possible to extend the lifespan of a GPU to 10+ years? If they bread down after 2-3 years at a 70%+ utilization rate?

Great question—focusing purely on physical durability under heavy 24/7 use. In data centers running intensive AI workloads at high utilization (60-70%+), GPUs typically last 1-3 years before thermal/electrical stress causes noticeable degradation or failure. At 95%+ sustained load, it's often closer to the 1-2 year range, though excellent cooling, monitoring (like NVIDIA DCGM), and maintenance can help push toward the upper end. They're engineered for reliability but aren't meant to run forever at full throttle.

I see, makes sense. Now I wonder, what are the general lifespans of GPUs? Irrespective of their economic value, let's say they're being used for as long as possible until they break or melt, supposing a very high utilization rate of 95%+. For how many years should such a chip run before it dies?

Great question! It's mostly the second point—**utilization**. GPUs aren't suddenly more durable physically. Instead, disaggregation lets older GPUs focus **only** on prefill (the memory-capacity-heavy phase), which plays to their strengths. They no longer have to handle the full inference load, especially the latency-sensitive decode phase. This keeps them economically viable for 10+ years instead of being replaced every 2–4 years as newer chips dominate the whole stack. Higher sustained value + lower obsolescence = much longer useful lifespan.

Solid explanation. I'd like to double click on expanding a GPUs lifespan. @grok why is it that GPUs suddenly life longer? - Because they now produce higher economic value than before? - They're utilized in a different way, running more tailored tasks like pre-fill instead of both (pre-fill and decode) which makes them life longer?

Yes, your summary nails the core idea. GPUs are versatile powerhouses optimized for the heavy parallel math in training, but inference has two phases with mismatched needs: prefill (memory-capacity heavy) and decode (low-latency token generation). Disaggregating them lets older GPUs handle prefill efficiently while specialized chips like Groq LPUs crush decode. The combo delivers higher effective token throughput and dramatically extends GPU lifespan/economics in inference clusters. Solid catch!

@HyperForgeAI Pair the more ambitious goal with an ownership coin and this project will thrive + onboard a bunch of loyal users with skin in the game.

We recently shared that we're pivoting toward a more ambitious goal - building an AI-first Three.js/WebGPU game engine. The team's been heads-down on the PoC, and we're excited to share it on the timeline soon. In the meantime, here's the new logo for HyperForge.

@sacha 🌱

metadao up ~70% from lows zcash up ~200% from lows last month was a great time to plant seeds. now is the time for patience

@bajpaiharsh244 Bad kids build.

This is so bullish, we already have a s-tier mobile wallet ecosystem with ZODL, and now we have another desktop wallet coming up. The zcash train only accelerates from here.

Introducing Vizor. A beautifully crafted @Zcash wallet for your desktop. Multi-account in one wallet. Keystone on day one. Open source. Available now on macOS: vizor.cash

@dogemos @vizorwallet Congrats bad boy. Good to see you and the team getting involved in the Zcash ecosystem. Best of luck Josh. Btw, does Vizor come with Ledger support for shielded ZEC? That would be amazing.

Our team has built wallets for years. @vizorwallet is that craft, applied to shielded ZEC. It's our first contribution to Zcash, but far from the last.

@dogemos 👀

🤫

@ApeDigit @rakka_sol @169rk @JosipVolarevic2 @omnipair @SebMontgomery I'm no longer in the video business. Thanks for the shout-out though. Glad you enjoyed my videos.

@rakka_sol @169rk @JosipVolarevic2 @omnipair @seppmos @SebMontgomery you should do a video on omnipair and break it down for us.

Was updating the @omnipair deck and every time I do it, I get reminded that Omnipair has ZERO competition. Which is awesome

@metaproph3t The pieces are coming together. Ownership alliance.

MetaDAO feels more like an ecosystem, in an L1-like sense, than I would have expected

@winniedydx Agreed, 100%

Been playing around with AI A LOT lately. Exciting times we live in truly.

Freedom money is making its moves. Seems like we're entering price discovery territory soon. Zillions!