“All men dream, but not equally. Those who dream by night in the dusty recesses of their minds, wake in the day to find that it was vanity: but the dreamers of the day are dangerous men, for they may act on their dreams with open eyes, to make them possible.” T.E.Lawrence
C8VC Daily Brief: AI's bottleneck is no longer intelligence. It's electricity. The market has only priced one of them.
— Google grew Cloud 63% to $20B and called itself "compute constrained." That's not a demand problem. That's a supply one.
— ~$640B of hyperscaler capex in 2026, climbing in 2027. None of it buys a smarter model. It buys power, land, cooling, silicon.
— One megawatt, leased to AI instead of hashing bitcoin: 5–7x the revenue. Same asset. The wrapper was never the value.
— Nuclear didn't get a climate bid this cycle. It got a compute bid from a buyer who doesn't care what a kilowatt costs, only that it's firm.
— Models are software. They compress toward zero. Power is physics. It doesn't.
The AI trade everyone owns is the model. The one still mispriced is the megawatt. When the scarce input is electrons, the re-rate goes to whoever already holds them.
$CIFR CEO Tyler Page teases future behind-the-meter on-site generation for Cipher Digital 👀
“Hyperscalers and neoclouds are looking for ways to get power faster. The easiest way in a lot of cases is tapping the pipeline by our property and generating our own electricity.”
I promised once our incubated companies Bitfury USA a.k.a $CIFR and Bitfury Canada a.k.a. $HUT would BOTH become DECACORNS I would make the tweet. WHAT a RIDE! Congrats to outstanding leadership by @rftylerpage and @ashergenoot and their respective teams. Next Stop - HECTACORN 🚀
DeCentralised AI @gonka_ai Time Has Come:
1) GPU Supply Crunch
2) Power Supply Crunch
3) Sovereign Mandates
4) Agentic AI Age
5) Massive Cost Advantage
Exactly! The sub-3s median latency and 100+ tokens/sec sustained throughput on frontier open-weight models like Kimi K2.6 and Qwen3-235B show that decentralized inference isn't a theoretical sandbox anymore —it is structurally fit for purpose for high-volume agent loops and asynchronous enterprise pipelines right now. In terms of the infra expect hardware evolution including ASICs ! They are coming.
Here's a quick summary:
https://t.co/eYlJXa0POf tested 2 frontier AI models (Kimi K2.6 & Qwen3-235B) on their decentralized network.
Kimi correctly knew “гонка” = Russian for “race.” Qwen hallucinated it as a Tamil board game. Neither knew they were running on Gonka.
Results: super low latency (1.26–2.83s), 67–100 tokens/sec, $0.10 per 1M tokens (20–750× cheaper than big-tech APIs), zero failures—even during a mainnet upgrade.
Easy OpenAI SDK integration, no Web3 hassle. Decentralized inference is scaling fast.
You can spin up your own testing containers, read the architecture logs, or grab 10M free developer onboarding tokens via the community framework gateway here:
• Access On-Ramp: https://t.co/UyswtryB3c
• Consensus Core Engine: https://t.co/HDxQMDjxi2
We asked TWO AI models running on @Gonka_ai decentralized network to describe Gonka.
Kimi K2.6 correctly identified the Russian etymology — гонка means "race" — which is genuinely the origin of the network's name.
Qwen3-235B confidently invented "a traditional Indian board game from Tamil regions."
One reasoned. One hallucinated. Neither knew about the decentralized AI network they were running on.
The infrastructure is scaling out faster than big-tech training-data cutoff timelines can index it.
Setting that irony aside — here's what 30 inference calls and $0.015 of automated testing actually revealed:
• Median latency on Qwen3-235B: 1.26s
• Median latency on Kimi K2.6: 2.83s
• Sustained Throughput: 67–100 tokens/sec
• Production Task Failures: 0 of 30
• Cost per 1M output tokens: $0.10 The same Kimi K2.6 model on Moonshot's own native API costs ~$2/M. That's 20× more for identical model outputs.
Versus other May 2026 frontier models:
• Gemini 3 Pro: ~$12/M (120× more)
• Claude Sonnet 4.6: $15/M (150× more)
• GPT-5: ~$40/M (400× more)
• Claude Opus 4.6: $75/M (750× more)
Total cost of the full 30-call benchmark: exactly one and a half cents ($0.015).
All run on the same day Gonka shipped its major v0.2.13 mainnet upgrade activating its live Ethereum cross-chain bridge. Zero dropouts or stalled inferences occurred during the entire core consensus upgrade window.
The developer experience requires zero Web3 friction: call the standard OpenAI SDK, swap your base_url parameter, drop in a network key, and ship. Done in 3 minutes.
The physical compute layer is scaling out globally much faster than the public distribution layer realizes it is operational.
GK DAILY BRIEF - we live in intersting times.
1/ G7 30-yr yields broke above post-Covid peak. The $50T safe-haven debt market is repricing for Iran-war inflation, not deflation.
2/ US Consumer Sentiment crashed to 44.8 — lowest in 74 years. Worse than Volcker '80, worse than '08, worse than COVID.
3/ 12-mo inflation expectations: 4.8%. The Fed's credibility anchor is unraveling in real-time.
4/ Sentiment -21% since Feb, -10% last month alone. Not a soft patch — consumer capitulation.
5/ Gerstner from Altimer: the AI shortage trade is a mirage. Hardware scarcity is artificially inflating low-quality commodity suppliers. Glaring mismatch vs resilient cash-generators like $NVDA. Mandate: own structural winners, not shortage beneficiaries.
6/ The trade no one's pricing right: power. AI demand + stagflation + grid constraints = utilities, nuclear, and miners-turned-HPC ($CIFR, $IREN) are the asymmetric infra bet. Hard assets + AI compute under one roof.
Stagflation regime + AI quality divergence + power scarcity. Position accordingly.
GK Daily Brief: Agentic AI Has Arrived. The Grid Hasn't.
Last night Jensen Huang said three words that change everything: "Demand has gone parabolic."
Here's what that actually means:
→ Nvidia just did $81.6B in revenue. Data center alone: $75.2B. Up 92% in a year.
→ Q2 guidance: $89–93B. Above every estimate on Wall Street
→ Agentic AI doesn't run in bursts. It runs 24/7. The compute demand curve doesn't flatten — it compounds → Hyperscalers confirmed $725B in AI capex for 2026. Six times 2022 levels !!!
→ ASML builds the machines that make the chips that power all of this. Their EUV tools take over a year to manufacture. You can't rush the choke point.
The market is still debating model benchmarks.
The real question is simpler:
Where does the power come from?
50% of US data center builds are already delayed — not because of capital or demand. Because the grid can't handle them. Transformer lead times: 5 years. Build time for a data center: 18 months.
Agentic AI running 24/7 doesn't just need more chips. It needs more watts. Every hour. Forever.
The chip race gets the headlines. The energy race gets the economics. Energy arbitrage IS the moat.
$NVDA $ASML $CEG $CIFR $IREN
GK Daily Brief: The Summit Happened. And Nothing Was Solved. ⚡
Stabilisation is not resolution. The market keeps confusing the two.
→ Brent pulling back from $120 to ~$109 on summit relief — but without SPR releases, diesel yield switching, and demand destruction all activating simultaneously, OECD diesel hits critical shortage levels by August, gasoline by October. The crude pullback is a relief valve, not a fix. It snaps back the moment any of those solvers disappoint
→ CLARITY Act Senate Banking Committee hearing completed — most serious legislative progress on crypto market structure in 2026. No vote timeline confirmed yet. But committee engagement is real. Watch for Senate floor scheduling. A vote = BTC breaks $82K on regulatory clarity alone
→ Warsh is now Fed Chair as of May 15. No press conference yet. He inherits CPI 3.8%, PPI 6.0%, real wages negative, and zero cuts priced. First Warsh speech or Fed minutes due this week. His view that AI boosts supply-side productivity and preference for trimmed mean PCE — which runs below core — gives him marginally more dovish cover than the headline numbers suggest
→ Pre-summit, the market priced a grand bargain. What it got was stabilisation — trade truce extended, Boeing jets and LNG purchases agreed, Xi told CEOs China will "open wider." No tariff reductions. Taiwan warning rattled markets. Fundamental issues unresolved. The tactical rally in Chinese equities and yuan is exactly what was expected. It doesn't change the structural picture
→ Nvidia earnings Wednesday May 20 — the most important corporate call of 2026. CEO just returned from Beijing without China chip access. Blackwell ramp, US data center demand, and China guidance will drive the print. At $5.4T market cap, every line of guidance moves markets. The China question is now the primary downside risk variable going into the call.
Energy arbitrage IS the moat — and while the summit bought time, the operators with secured power don't need diplomacy to win. The grid constraint compounds regardless of what happens in Beijing.
$NVDA $CEG $CIFR $IREN $PWR $MU $VST $VRT