Welcome to #102 of the AI edge.
The week started with cautious optimism as BTC bounced off 70k. That didn't last. By Friday we were staring at 60k and praying for the red candles to stop.
Saylor added insult to injury, selling again this week. The man's strategy (pun intended) at this point is buy every top, sell every bottom. Truly one of us.
On the AI side, all eyes were on Taiwan as Computex pulled in nearly every major semiconductor company on the planet. Nvidia and Intel both showed up but kept the data center announcements in the drawer, focusing on new PC silicon instead. Edge computing was the big narrative.
Jensen also made his pick for the next trillion-dollar company: Marvell. Not the superhero franchise. This one makes chips. (Full profile on our Tessara terminal if you're curious.)
The other big headline was Anthropic confirming it's confidentially filed for an IPO. SpaceX first, now Anthropic. Two trillion-dollar companies potentially hitting public markets in the same month. Equity markets are about to get quite interesting!
With that, let's get into this week's edition.
The Big Story: 100 Billion Parameters, No Datacenter
Back in March, we covered Covenant-72B, the largest decentralized training run anyone had pulled off at the time. Shortly after, we also covered the fallout when the project's founder dumped roughly $10M in TAO and things fell apart on CT.
Bittensor took the reputational hit. The network kept shipping anyway. This week, a different team on a different subnet blew past that benchmark entirely.
IOTA, on Bittensor's Subnet 9, just pretrained a 100-billion-parameter model across 48 individual A100 GPUs scattered around the open internet, with each GPU sat in a different location, coordinated into a single training system.
What makes this different from Covenant is that Covenant required every machine in the network to hold a full copy of the model, so the biggest thing you could train was limited by the smallest machine in the room. And joining meant running eight B200 GPUs at $50 an hour.
IOTA took a harder path. They split the model itself into pieces, one per GPU, and each participant handles a slice before passing it along. Model size grows with the network. More peers, bigger model.
The Wedge
The run averaged 30% model FLOP utilization, which measures how much of a GPU's raw power actually goes toward useful training. Frontier labs with purpose-built clusters typically hit 38-50%. Reaching 30% on scattered commodity hardware over regular internet puts IOTA at roughly 65% the speed of a top-tier datacenter setup. For what they're working with, that's a serious result.
Cost per participant: ~$1.4 an hour. A full replica ran at $20/hour compared to $50 for Covenant's equivalent. That gap widens once spot pricing and consumer GPUs enter the picture.
If this scales, it's the first credible challenge to the assumption that frontier training requires a frontier datacenter.
The Fine Print
The run trained on 1.1 billion tokens over two days before the team stopped it for cost, so there's no finished model to show and no benchmarks to stack against frontier outputs. Taking this system and producing something competitive end-to-end is the real test, and that hasn't happened yet.
The hardware was still cloud-provisioned A100s sitting in US datacenters, which is a long way from the end vision of mixed hardware from untrusted participants on consumer GPUs. All of that is on the roadmap but none of it has been proven.
It was 72 billion a few months ago. Now it's 100 billion. A trillion-parameter decentralized training run on Bittensor doesn't feel that far away.
Constraints Watch: The Steel Inside The Transformer Shortage
Everyone agrees power is becoming a big bottleneck for the AI buildout now. Generation, grid capacity, years of permitting. All real, all well covered.
The part that gets less airtime is what happens once the power exists. Moving it from the grid into a data center requires a large power transformer, or LPT, a custom-built machine north of 100 tons that converts voltage between the plant and the substation feeding a campus. Each one is engineered for a single site. Nothing ships off a shelf. Without one, every megawatt a hyperscaler signed for never leaves the grid.
Order books at the transformer makers are the deepest in decades and they're still falling behind. Lead times have tripled since 2021. An estimated half of US data center projects lined up for 2026 may slip or fall through entirely, all on power delivery.
From GE Vernova’s (a major Transformer Manufacturer) Earnings Analysis on Tessara
But the deeper story is what the transformer is actually made of.
Its core is wound from grain-oriented electrical steel, or GOES, a specialty alloy that barely leaks energy as heat. There's no substitute at high voltage, and qualifying a new production line takes years and hundreds of millions of dollars. So the global supplier base is a handful of names: Baowu, Nippon Steel, POSCO, Thyssenkrupp, and Cleveland-Cliffs.
Inside the US, that list is one name long. Cleveland-Cliffs runs the only domestic GOES line, out of a single works in Pennsylvania. No second source. No backup.
And the US still brings in more than 80% of its large transformers from overseas while the demand is overwhelmingly American. That mismatch is starting to matter.
How much it matters, what's moving in Cliffs' direction, and why the market may be mispricing it. That's all in this week's Chokepoint. (our AI infra newsletter)
Virtuals Protocol integrated Venice for private, uncensored inference on Base and is deploying $400K in credits to bootstrap agent builders.
Walrus introduced Walrus Memory, a portable, user-controlled memory layer that lets AI agents carry persistent context across apps without platform lock-in.
OpenGradient launched OpenGradient Chat, a privacy-first AI app offering frontier models with encrypted, unreadable user prompts.
Wayfinder unveiled Paths, a plugin marketplace for onchain agents where users can install or publish skills and trading strategies secured by collateral.
Quasar began training a new 10B long-context model, leveraging Teutonic (Bittensor SN3) for decentralized pretraining while Quasar focuses on distillation and architecture refinement, advancing distributed LLM development at scale.
Trishool (SN23) joined the Google for Startups Web3 Program, securing up to $200K in cloud credits to accelerate AI safety and alignment efforts.
Virtuals is migrating its $700M+ VIRTUAL token to Chainlink CCIP, upgrading cross-chain security for agent infrastructure.
Endure Network (SN30) launched a decentralized risk intelligence network on Bittensor, with its native TAO lending market “Forge” coming soon.
🔥 Our Weekly Top Tweets
#1 Zcash Crashes 50% After Claude Identifies Critical Bug
$ZEC plunged over 50% after a long-undetected bug in its Orchard privacy pool was revealed by using Claude Opus 4.8. While the flaw is patched, Zcash’s shielded design makes it impossible to fully verify whether counterfeit coins were minted before the fix and the uncertainty alone triggered a ~$5B market cap wipeout.
#2 Bankr Is Becoming the Default Launchpad on Base
Bankr just hit 90.5% daily launchpad market share with an 80.3% 7-day average, both all-time highs on Base. It is beating some big names in agentic launchpads like Virtuals, Clanker, etc.
Cheers,
Teng Yan & Arvind
Quick recap: I launched The Chokepoint, a new weekly newsletter on the AI supply chain. Free, every Tuesday on Tessara Research.
