Capital found a workaround for every constraint this week except one.
The cap table absorbed its supply chain. The cluster outgrew its building. A once and future competitor leveraged acquihire attention to reenter the game. An API scaled multi-model, multi-modal access. But permits?
Refusals and moratoriums.
Tough to build when the ground says no.
I'm Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
I’m actively tracking pricing sentiment across the industry for new content. For that to work, I need your help with a quick question.
Where is GPU compute pricing heading over the next 6 months?
p.s. Respondents will get priority access to reports as they’re created…
Anthropic Raises $65 Billion and Launches Opus 4.8
Revenue grew $17 billion in two months. The supply chain bought in.
Anthropic raised their Series H this week. Altimeter, Dragoneer, Greenoaks, and Sequoia led the round. Capital Group, Coatue, D1, GIC, ICONIQ, and XN co-led, with Temasek and MGX also participating. The $65 billion headline includes $15 billion of previously committed hyperscaler capacity, including $5 billion from Amazon, alongside ongoing capacity deals with Google/Broadcom (5GW TPU) and SpaceX (Colossus 1 and 2). Claude Opus 4.8 launched the same day with dynamic workflows running hundreds of parallel subagents in Claude Code and fast-mode pricing three times cheaper than the prior generation. AWS remains primary cloud and training partner.
Why this matters:
Anthropic names Micron, Samsung, and SK hynix separately as "strategic infrastructure partners" on memory.
If this ends up looking anything like OpenAI’s previous cornering of memory supply, it’s good for Anthropic and bad for absolutely everyone else. From a CapEx and consumer standpoint, at least.
The $965 billion post-money valuation prices Anthropic at roughly 20x current run-rate revenue. OpenAI's most recent round priced at approximately 15x. Big multiples in the run up to a potential IPO.
Groq Returns as an Inference Neocloud After Licensing Its LPU IP to NVIDIA
Groq is reportedly raising fresh capital to re-enter the infrastructure layer it sold out of last December.
According to Axios, Groq is reportedly raising approximately $650 million from existing investors to fund Groq 2.0, led by company veterans Adam Winter (CEO) and Matt Eng (CFO). Disruptive and Infinitum are backstopping the round. The raise follows NVIDIA's $20 billion licensing deal for Groq's LPU stack in December (Issue #80), which paid out original Groq investors and triggered senior team departures. Groq 2.0 pivots the company back into the infrastructure layer it exited, this time as an AI inference neocloud rather than a chip designer.
Why this matters:
Groq 2.0's leadership (Adam Winter, Matt Eng) are company veterans, retaining the operational knowledge of LPU deployment after senior team departures tied to the NVIDIA deal (Issue #80).
Groq 2.0 enters a neocloud field that includes CoreWeave, Lambda, Nebius, IREN, and the Blackstone/Google TPU venture (Issue #107).
Given the ongoing market shift from training to inference, and Cerebras’ recent IPO success, Groq 2.0 may yet have its moment in the sun.
IREN Signs $1.6 Billion Dell Deal for Blackwell Systems at Childress
When is a target not a target? Asking for a friend.
IREN announced a $1.6 billion purchase agreement with Dell for air-cooled Blackwell systems to service its five-year, $3.4 billion managed services AI cloud contract. Systems deploy at IREN's Childress, Texas campus, which operates on a 750MW power footprint shared with its Bitcoin mining operations. Commissioning is targeted for early 2027, with ARR rising from $3.7 billion to $4.4 billion.
Why this matters:
The target breaks down as $1.9 billion from the Microsoft contract and $0.7 billion from the AI cloud contract. The remaining $1.8 billion is from planned GPU deployments at British Columbia and Childress based on internal assumptions for utilisation and pricing.
Internal projections for unsigned managed services revenue are common practice in the neocloud category; CoreWeave, Nebius, and Lambda all carry contracted-to-projected gaps in their public disclosures.
Even so, $1.8 billion is a big number to target on assumptions alone. Maybe there’s something they aren’t telling us? Yet, at least.
OpenRouter Raises $113 Million from Alphabet, NVIDIA, and the Data Platform Stack
Why have multiple API’s when you can have one?
OpenRouter has raised $113 million in Series B. CapitalG (Alphabet) led the round, with participation from NVentures (NVIDIA), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, AMP PBC, and Pace Capital, alongside existing backers Andreessen Horowitz and Menlo Ventures. The round values OpenRouter at approximately $1.3 billion, more than double its valuation 12 months ago.
Why this matters:
For those unfamailiar, OpenRouter provides a unified API gateway that routes inference requests from applications to whichever model and provider best fits each call. To put it simply, OpenRouter simplifies the API experience.
Weekly tokens routed grew 5x in six months, from 5 trillion to 25 trillion. With multi-model production deployment now the dominant share of enterprise AI inference traffic, OpenRouter is poised to capture a share of the growing enterprise market.
The platforms enterprises already run AI workloads on (Alphabet via CapitalG, NVIDIA via NVentures, ServiceNow, MongoDB, Snowflake, Databricks) all now hold equity in the layer that routes inference between applications and models.
FuriosaAI Partners With Broadcom to Build a 2nm Inference Platform for Hyperscale
Broadcom has added another partner to their roster.
FuriosaAI is partnering with the custom chip manufacturer on their third-generation inference accelerator. The chip will have a 2nm compute die, HBM4/4E memory, multi-die chiplet packaging, and Broadcom Ethernet for rack-scale clusters. Gen 3 builds on RNGD, FuriosaAI's second-generation chip in mass production at TSMC 5nm, validated by Samsung SDS and LG AI Research on production Korean-language inference workloads. FuriosaAI has raised $250 million to date.
Why this matters:
Gen 3 ships as a chip-plus-networking platform with rack-scale interconnect specifications integrated into the silicon design, positioning it as an Ethernet-based alternative to NVIDIA NVLink topologies.
H1 2028 sampling places Gen 3 in the same window as NVIDIA's post-Rubin roadmap.
That means we have Cerebras (Issue #106) targeting wafer-scale, Fractile (Issue #105) targeting SRAM-native, and now FuriosaAI targeting 2nm-plus-Ethernet, each addressing the same inference market with a different architectural bet.
Google Details Virgo: 1 Million TPUs Across Sites in One Training Cluster
The single-site power ceiling just stopped being a ceiling.
Google detailed Virgo Network, its scale-out data centre fabric for the AI era. A single Virgo fabric links 134,000 TPU 8t chips with 47 petabits per second of non-blocking bisectional bandwidth, delivers 4x the bandwidth per accelerator over the previous generation, and reduces unloaded fabric latency by 40%. With Pathways and JAX, the fabric scales to near-linear performance across a million TPU 8t chips in a single logical cluster, spanning multiple data centres. The architecture decouples scale-up, east-west scale-out, and Jupiter frontend traffic into three independent fabric domains. The "campus as a computer" philosophy positions multi-site training as a default, not a fallback.
Why this matters:
Virgo distributes training across multiple data centres at smaller individual power footprints, removing the gigawatt-class single-site requirement as a blocker for frontier-scale training.
Virgo Network is the connective tissue of Google's AI Hypercomputer stack. The Blackstone/Google TPU neocloud (Issue #107) is the first commercial buyer of capacity built on this fabric.
Multi-site training shifts site selection from single-site power capacity to inter-site connectivity. Fibre routes, dark fibre availability, and bandwidth lease economics become primary inputs.
Minneapolis Approves a Six-Month Moratorium on Large Data Centres
Capital wants to build. Minneapolis just said no.
Minneapolis City Council approved the moratorium, the latest in a series of construction freezes across Minnesota. The 350,000 square foot threshold sits below the average footprint of new hyperscale builds, which typically span 500,000 to 1,000,000 square feet. Council documentation cited water use, transformer load, and tax abatement structures as the basis for the freeze. Minnesota hosts roughly 1.2GW of operational data centre capacity, per state utility filings, with further capacity in active development.
Why this matters:
Google and Nebius announced 2.4GW of combined capacity in Missouri last week (Issue #107) while PJM secured curtailment rights in Virginia. Some states are bidding for the buildout. Others are stopping it at the door.
New applications during the freeze are paused, not denied; developer commitments shift to other Minneapolis suburbs or out of state.
Permitting risk is now a material line item in site selection underwriting. In the Twin Cities, at least.
The Rundown
The AI infrastructure stack is being rewired around constraints rather than within them.
When a single site can't hold the cluster, the cluster spreads across sites. When the silicon incumbent buys out the chip company, the chip company takes the money and re-enters the field as a cloud. When the customer outgrows its capital structure, it borrows from the suppliers that built the compute. When enterprises need to route inference across many models, the unified API above them gets funded by every platform they already run on.
Necessity is the mother of invention.
But each adaptation costs something.
Multi-site fabrics add fibre dependence. Non-NVIDIA silicon adds integration risk. Supply-chain capital ties customer growth to supplier outcomes. Unified routing adds gateway dependency.
Risks can always be managed.
But only if you get the permits needed to build the shed in which all of the infrastructure lives.
See you next week.