IA_817

RT isn’t an endorsement and doesn’t reflect the views of my employer; often, it’s to revisit a post

[Ornn @OrnnExchange]

https://t.co/3LiNa0iQsC

[Bill Ackman @BillAckman]

An important read on the hyperscalers and why we are bullish.

[zerohedge @zerohedge]

Goldman Delta 1 head:

"The most interesting thing I read over the weekend came from Brian Armstrong at Coinbase. The accompanying chart showed AI spend has almost halved while token usage continues to grow exponentially. As Armstrong put it, “The goal isn’t to suppress usage. It’s to build the infrastructure that makes exponential growth sustainable.” The key wasn’t using less AI, but routing. Simple tasks are automatically sent to cheaper local or open-weight models, while frontier models are reserved for genuinely difficult reasoning. As he put it, “Ultimately, humans shouldn’t be choosing models. AI can automate this task.” I think this is one of the clearest examples yet of inference cost deflation. Some enterprises will build on-prem infrastructure, but many will simply route workloads to lower cost providers. The common denominator is that useful output per dollar keeps rising. Ironically, if companies discover they can achieve the same output for half the AI spend, the first instinct may not be to accelerate capex… it may be to pause and optimize."

[Keith Rabois @rabois]

Worth reading carefully.

[Michael McNair @michaeljmcnair]

The doomer argument isn’t that datacenter profitability is bad right now.

It’s that the profitability is being underwritten by model companies buying compute at extreme prices while running unsustainable economics themselves.

The key variables in the model (GPU-hour pricing and utilization) depend on the model companies continuing to raise capital and buy compute at prices that might not be supported by end customer profitability.

If investors keep funding model company cash burn, the datacenters look great. If they decide model company spending is uneconomic, then compute demand gets rationalized and datacenter pricing/utilization fall with it.

Right now the datacenters can earn great returns because the cash burn is sitting one layer downstream.

[Carnivore Aurelius ©🥩 ☀️🦙 @AlpacaAurelius]

This analysis of over 300,00 people showed that sunscreen has NO effect on reducing skin cancer.

Other studies even show sunscreen is correlated with higher rates of skin cancer

Sunscreen is a massive scam.

[Rohan Paul @rohanpaul_ai]

👀 This is from an ex-Meta PM.

[andrew arruda @andrewarruda]

this is a really smart move

[Molly O’Shea @MollySOShea]

NEW: The Death of Spreadsheet Investing. Inside Benchmark's New AI Playbook

Everett Randle (@EverettRandle), GP at @benchmark

Why every golden rule of SaaS just got inverted & how AI is rewriting venture investing:

› You can now pass $1B+ in revenue with unproven unit economics..
› Late-stage companies can have much higher upside than a Series C

Benchmark's AI Bets: Cerebras, Manus, Mercor, Sierra, Legora, Fireworks, Exa, Starcloud, Gumloop..

We cover:
› The Disorientation: Scale No Longer De-Risks a Company
› What Happened to the Golden Rules?
› AI Inverted Nearly Every Framework
› New Taxonomy: P x Q x M
› Inference Is the Demand Engine
› Tokenomics: Frontier vs Open Source
› The Liquidity Shock: Reframing the Returns Math
› Benchmark's Model: Founders Over Themes

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:00) Everett Randle, General Partner at Benchmark
(00:58) Coming off Benchmark's AGM
(04:05) The Golden rules of Investing are all gone
(08:58) Who actually has a handle on AI Economics?
(12:57) Why Benchmark bets on Founders, not Categories
(15:31) Brad Gerstner's "Age of Inference" Thesis
(19:07) The most important shift since the Cloud
(23:55) Inside Gumloop's AI automation canvas
(26:40) The Token Maxing problem nobody's solving
(27:25) Ev's "Mom Test" for frontier AI
(31:33) What happens when frontier models get too cheap
(34:44) Inside the new funding playbook
(40:31) Venture Capital became a product, not a firm
(42:35) The Biggest IPO wave Wall Street's ever seen
(49:08) The secret behind Benchmark's wildly diverse bets
(52:54) The mentors who shaped him

[Chris Brown @almostcmb]

Was talking to a few people about this over the weekend. I think it's reasonable to ask if this puts the category out of contention for early stage (pre-seed/seed) capital tasked with being non-consensus. People will fall on both sides of that question. I don't think it does.

A few reasons:

> The lazy answer is "but it's going to be such a big market!!" Which falls victim to the same logic @0xsmac laid out in a tweet yesterday about convincing ourselves there's a pot of gold at the end of all parabolic curves. The better version of this answer is that the forces that would propel the industry towards that outcome are inherently venture-compatible: scaling laws, genuine R&D, novel technical milestones, data network effects. We're like ~24 months into those efforts and they are significantly more diffused across teams and companies than AI was in 2020 when OAI first told everyone they had figured out how to make parameters, data, and compute drive predictable model improvement.

> I think it's reasonable to ask whether $16b is *actually* a big number. That is similar in size to categories like healthcare, fintech, bio, and non-AI SaaS, which by comparison are smaller markets with known ceilings and are mostly all betting on replacement cycles to succeed. But also, there is an enormous amount of concentration in this number. I'm doing this off the top of my head, but I think like ~50% of that Q1 total is accounted for by rounds into Shield/Skild/Apptronik/Bedrock/Neura/etc.

> And that gets us to the final point. Capital flowing into the category to date has been pretty unimaginative. There is 100% froth. But it's concentrated at the top of the market in a relative small set of "trades" - humanoids, defense w/ contract revenue, and labs. Physical AI is not a monolith, and unlike AI, the froth has not bled into all companies within the category (yet). In my own pitch experience, I've seen some genuinely novel stuff across unique embodiment form factors, software infra, dataset curation, developer tooling, and vertical-specific industry approaches. Many of these are having a hard time raising money because they aren't labs and don't have a $100m contract with the DoD. Investors really don't know how to allocate capital in this category yet outside a handful of archetypes.

[Brett Harrison @BrettHarrison]

A Guide to the US Financialization of Compute

We frequently receive questions about compute trading from neoclouds, GPU-as-a-service providers, data center operators, training and inference companies, energy companies, HFTs, brokerages, investment banks, FCMs, CTAs, RIAs, ETF issuers, VCs, and others. Here's a current summary of the market structure and participants:

Three main participants in the US compute derivatives universe:

1. CFTC-regulated derivatives exchanges (DCMs)

Definition
• Exchanges where CFTC-regulated futures, options on futures, and swaps can be legally traded in the US.
• Designs, certifies, and lists derivatives contracts. Engages with third-party index providers for cash-settled derivatives’ underlying settlement prices.
• Facilitates capital formation and investment in new US commodity, currency, and energy products.
• Responsible for market monitoring, position limits, circuit breakers, recordkeeping, and protection against market manipulation.

Participants
• American Innovation Exchange: Architect acquired a DCM this year to launch the first AI-industry-dedicated futures exchange in the US. Going live soon with listed compute futures and options, with index data from Compute Desk.
• CME: the largest US futures exchange by volume, concentrated in stock indexes, rates, and agriculture. In an exclusive agreement with data provider Silicon Data to list compute futures later this year.
• ICE Futures US: the second-largest US futures exchange (run by NYSE's parent), concentrated in energy/power and soft agricultural derivatives. ICE announced intent to list compute futures in an exclusive agreement with data provider Ornn.

(pending regulatory review)

Role in compute futures/options
• Create derivatives contracts that allow commercial compute consumers, compute producers, financial firms, and individuals to hedge and speculate on the price of compute for different GPU types.
• Build a broad liquidity profile to create price discovery across the futures expiry curve.
• Facilitate sufficient liquidity and volume for the creation and redemption of compute ETFs, currently registered by six ETF issuers.

2. Compute index providers

Definition
Independent third parties that combine rental price offers and private transactions into single values representing the cost of compute per accelerator.

Participants
• Compute Desk, Silicon Data, Ornn, SemiAnalysis.
• Free aggregators (not formal index calculators): United Compute, AI Multiple, GPU Lease Index, CloudePrice_net, Thunder Compute.

Role in compute futures/options
• Standardize pricing data across the range of GPU manufacturers, SKUs, configurations, and geographies to build a useful index for hedging by commercial consumers and producers of compute.
• Build non-manipulable, IOSCO-compliant benchmarks for use in CFTC-regulated cash-settled futures contracts on DCMs.

3. Spot/forward compute capacity platforms

Definition
• Marketplaces that match customers seeking short- and long-term capacity with the neoclouds and GPU-as-a-service providers that make delivery.

Participants
• Nvidia DGX Cloud Lepton, Compute Desk, Compute Exchange, Vast_ai, Andromeda, VoltagePark, HydraHost, RunPod, Ornn, SF Compute, Shadeform, Spheron, Hyperbolic, SaladCloud, Prime Intellect, Clore_ai, Cudo Compute, Akash Network, Digital Ocean, Aethir.

Role in compute futures/options
• Collect and normalize raw GPU pricing data for index providers.
• Aggregate fragmented physical supply/demand and establish compute grades as precursor to development of physically-settled exchange-traded futures.
• Create transparency in a market where suppliers may prefer opaque pricing power.
————————————————————
Architect’s thesis is that compute will mature as a US exchange-traded asset class very quickly. We're excited to compete with incumbent exchanges, support index providers, provide ETF liquidity, and partner with capacity platforms to augment our cash-settled futures markets with physical delivery capabilities.

[Andrew Côté @Andercot]

Nuclear Batteries are completely slept on

100x-1000x energy density of Li-Ion, decades long duration of stable power output

Rovers on the dark side of the Moon

Deep space satellites far past the Belt

Drones lying in wait for years at the bottom of the ocean

[Noah Smith 🐇🇺🇸🇺🇦🇹🇼 @Noahpinion]

"AI is a bubble" is looking less and less likely by the day

[Dylan Patel @dylan522p]

People don't quite understand how many behind the meter power generation assets are being built for datacenters because the US Grid sucks, despite the higher cost and complexity.

[Spencer Applebaum @SpencerApplebau]

1/ Today, @multicoin publishes our HYPE analysis and valuation

HYPE is now one of our largest liquid fund positions and we've been accumulating aggressively since February

Full report and disclosures in link. Link here -

https://t.co/tA2kwWbG9Q

[Elon Musk @elonmusk]

Lee Kuan Yew was a genius

[Ibrahim Ajami @IbrahimAjami]

As investors in Baseten competitors (the inference market is simply too big and important to ignore), we as investors broadly usually just hype our own portfolio.

But @apoorv03’s latest post from Altimeter is genuinely excellent—deep, thoughtful, and high-signal.

Well done

[Ori Eldarov @leveredvlad]

https://t.co/RNnHiISayr

[RJC @RJCcapital]

This is my equity analyst from now on for these types of post just follow this guy put notif on

[Gregory Blotnick @gregoryblotnick]

"Tiger Cub Investment Framework" -- taught by John Griffin in his Security Analysis class at Darden & CBS decades ago