Beyond the compute boom: BlackRock, Bain, AMD and Venice — Wed 30 Sep

Good morning!

Today's Kaupr Daily is about AI and compute: how computing power is starting to be treated as an asset, and how demand is shifting from training models to running them. We also look at what the build-out has to earn, and at where AI is heading next, into the physical world and onto decentralised networks.

Some stories in Kaupr Daily:

💎 BlackRock sees compute becoming a tradable asset
💎 AMD buys Fei-Fei Li's World Labs for $8.2 billion to prepare for physical AI
💎 Swedish-founded Modal Labs nears a $15.75 billion valuation
💎 Bain: AI needs $6 trillion a year by 2031 to justify the build-out
💎 Venice Token is up nearly 1,800% as privacy AI gains users

— Morten

We covered BlackRock's AI report last week; today we return to its compute angle. BlackRock envisions standardised contracts on computing power, including exchange-traded futures, that could be tokenised, pledged as collateral and settled onchain. It also cautions that the market is illiquid and that chip generations and electricity prices differ widely.

Why it matters: If compute can be priced, hedged and pledged like other commodities, the AI build-out gets a financial market of its own.

Source: BlackRock sees AI as a new driver for digital assets — Kaupr

On the Milk Road podcast, analyst Vincent argues that cheaper open-weight and specialised models are taking token share from OpenAI and Anthropic. Because those labs are major buyers of compute, a slowdown could hit AI infrastructure. His longer-term case: AI agents' inference needs lift demand, and stablecoin rails can serve agents that transact continuously.

Why it matters: The first phase of the AI build-out rewarded whoever owned the bottleneck. The next may favour those who turn AI into everyday use, including payments between machines.

Source: The AI Compute Trade Is at Risk - Here's What Could Win Next — Milk Road

Grayscale argues demand for compute and supply of it are drifting apart. AI agents can use 5–50 times more compute than a chatbot conversation, while power, data centres and chips take years to build. Data centres may account for about half of US electricity-demand growth through 2030, and grid connections can take over five years.

Why it matters: In Grayscale's view, value will accrue to those who already own energised compute capacity: power producers, data centre operators and AI clouds.

Source: Why Compute Is the New Scarce Digital Asset — Grayscale (Source: company research)

Inference, running trained models, now accounts for two-thirds of AI compute demand, up from one-third in 2023. Spending on inference is expected to pass training for the first time this year. The nine largest cloud providers are projected to invest $830 billion in 2026, almost 80% more than last year.

Why it matters: Training is a one-off cost; inference runs every time a model is used. That moves the bottlenecks to memory, networking, power and cooling.

Source: AI Compute Scales Tilt Toward Inference — BigGo Finance

Inference provider Modal Labs is close to raising $750 million led by Accel at a $15.75 billion valuation, more than triple its value four months ago, TechCrunch reports, citing a source. CEO Erik Bernhardsson is Swedish and helped build Spotify's recommendation system. Rival Baseten is reportedly being valued at $26 billion.

Why it matters: Revenue at inference providers is growing fast, yet margins stay thin because compute remains expensive.

Source: Inference provider Modal Labs closing in on $750M round at $15.75B valuation — TechCrunch

Neocloud General Compute has signed a multi-year deal to deploy Cerebras chips, starting with coding agents that make many sequential inference calls. The companies say lenders are increasingly funding such hardware purchases, so providers can offer fast inference without carrying the chips on their own balance sheets. General Compute has a $400 million debt facility.

Why it matters: When AI hardware can be financed like other equipment, compute starts behaving like a financial asset.

Source: General Compute Selects Cerebras to Bring Ultra-Fast Inference to Agentic Coding — Cerebras (Source: press release)

Bain says the AI industry needs $6 trillion in annual revenue by 2031 to justify data-centre spending, and that infrastructure is being built well ahead of demand. Funding it sustainably would require adding about 1% to global GDP growth each year. Existing services may bring $1.8 trillion, leaving $4.2 trillion to be created.

Why it matters: Shortages of power, transformers and water, and local opposition, are already slowing data-centre projects. Whether the pace holds depends on new revenue arriving as fast as the build-out.

Source: AI Faces $6 Trillion Test to Justify Data Centers, Bain Says — Bloomberg

AMD has agreed to buy World Labs, the AI lab led by Fei-Fei Li, in an all-stock deal worth about $8.2 billion. World Labs builds models that generate and simulate 3D environments, and technology for robotic learning. AMD says compute needs grow more varied as AI expands into robotics and simulation.

Why it matters: A chipmaker is buying a model lab to learn what the next workloads will demand, and it expects them to come from robotics and simulation.

Source: AMD to Acquire World Labs to Advance the Future of AI Compute — AMD (Source: press release)

Venice AI, a decentralised, privacy-focused AI platform co-founded by Erik Voorhees, raised $65 million in July at a $1 billion valuation. Its token, VVV, is up nearly 1,800% this year. Staking VVV gives users a share of inference capacity or mints DIEM, a tradable compute credit. Most tokens are held by a few large holders.

Why it matters: Venice turns compute into something users can hold and trade. Part of the rally came from liquidations of leveraged positions, which the source flags as a risk.

Source: This AI Token Is Up Nearly 1,800% For the Year — The Motley Fool

On Nosana, a decentralised GPU marketplace, developers run agent frameworks such as ElizaOS and Mastra, and payment settles through smart contracts. An agent's wallet works like an open credential: whoever holds the key can spend the full balance, irreversibly. Solana supports capped, revocable spending permissions, but limits on purpose and expiry are not yet built.

Why it matters: When machines spend money on their own, mandate becomes a security question. A leaked key can be rotated; a confirmed payment cannot be reversed.

Source: Who Holds the Authority When AI Agents Pay for Their Own GPU Compute? — HackerNoon (Source: opinion piece)

Kaupr has launched Kaupr Events, a short newsletter about events in onchain finance: our own live events, curated picks from the Nordics and Baltics, and recordings of what you missed. It replaces our Luma calendar as the home for our events.

First up is How to invest in the tokenization trends, a joint event with Virtune on Wednesday 7 October, 11:00–12:30 CET, live online. Virtune's CEO and co-founder Christopher Kock and Chief Sales Officer Andreas Severin discuss what is driving tokenization, which platforms the financial industry is building on, and how investors can gain exposure through regulated ETPs. Moderated by Morten Myrstad.

Read the first issue → · Register for the event →

Wishing you a great Wednesday — and welcome back tomorrow morning for the next edition of Kaupr Today.

Best regards
Morten Myrstad
Founder & Editor