OPEN SIGNAL
Briefings ·

Signal Briefing: June 2, 2026

Anthropic's confidential IPO filing, a reported 250GW data center capacity gap, and SK Hynix's commitment to double memory output by 2031 together define a week when AI infrastructure economics forced themselves onto Wall Street, Capitol Hill, and the foundry floor simultaneously.

Anthropic Confidentially Files for IPO, Topping OpenAI’s Valuation

Anthropic has filed confidentially with the U.S. SEC to go public, with share count and price not yet disclosed, according to Data Center Dynamics. The Register notes the company now tops OpenAI in valuation and is beating Altman to the public markets, racing OpenAI and SpaceX to Wall Street in what may become the defining capital-markets moment of the AI cycle.

Why this matters. A public Anthropic creates the first directly comparable revenue and margin data point for frontier model providers, forcing the entire sector — hyperscalers, inference clouds, and enterprise buyers — to price AI model economics against a real income statement rather than private-round speculation. The S-1 disclosures, once public, will set expectations for the cost structure of compute-heavy model businesses.

Confidence: high — Multiple outlets confirming a primary SEC filing event; no price or timeline figures cited beyond public record.


Industry Report Flags 250GW Data Center Capacity Gap Despite Record Investment

A new report cited by Data Center Dynamics finds that AI data center demand is “larger than we’re prepared for,” with an additional 250GW of capacity required under what it calls an “aggressive but still plausible” scenario — even as the industry is already making what it describes as “existential investment.”

Why this matters. The framing signals that announced hyperscaler and sovereign capex commitments — running at roughly $50–100B/year per FY24–25 10-K disclosures across the major cloud providers — are not closing the gap; they are simply keeping pace with the acceleration of the baseline. Every megawatt of unbuilt capacity is a constraint on inference throughput and, ultimately, on the cost curve that determines AI’s commercial viability at scale.

Confidence: medium — Single trade report; the 250GW figure is from an analyst scenario model, not a primary operator disclosure.


SK Hynix to Double Memory Wafer Output Within Five Years, Sees Shortage Through 2030

SK Group chairman Chey Tae-won told reporters at Computex on June 2nd that SK Hynix will double its memory wafer capacity within five years, per Tom’s Hardware. The company projects that an AI-driven memory shortage will persist until at least 2030, underscoring that supply cannot respond to demand on a two-year cycle.

Why this matters. HBM and high-bandwidth DRAM are among the tightest constraints in the AI accelerator supply chain today; SK Hynix is the dominant HBM supplier for Nvidia’s H- and B-series GPUs. A five-year timeline to double output confirms that memory will remain a structural bottleneck through at least the mid-2030 compute buildout, with pricing implications for every hyperscaler and neocloud buying accelerator clusters at scale.

Confidence: high — Primary disclosure from SK Group chairman at a named public event (Computex, June 2, 2026).


Intel and SambaNova Deploy Disaggregated Inference at Rack Scale; Vista Equity Backs Inference Neocloud on the Same Stack

Intel and SambaNova have landed their first customer for a disaggregated inference architecture that crams 36,864 CPU cores into a 100kW rack targeting agentic AI workloads, per The Register. In a related move, Vista Equity Partners has launched Vector Core Compute, an inference-focused “agentic neocloud” built on SambaNova and Intel compute, according to Data Center Dynamics, citing demand it observed inside its portfolio of enterprise software companies.

Why this matters. Both moves represent the same thesis from different angles: that agentic AI inference workloads — high-concurrency, latency-sensitive, CPU-memory bound — do not fit neatly onto GPU-centric training clusters, and that an alternative inference compute market is now large enough to attract both venture-backed neoclouds and enterprise chipmaker partnerships. If the disaggregated CPU inference model scales, it introduces a credible non-Nvidia option into the inference cost stack for the first time.

Confidence: medium — Two corroborating sources on the same stack; first-customer claim for the Intel/SambaNova blueprint is from a single trade outlet.


Ohio Pauses Data Center Tax Breaks After Finding Itself in the “Billion Dollar Losers’ Club”

Ohio has suspended its data center tax incentive program after determining it had ceded far more revenue than anticipated, with The Register reporting the state found itself in what officials described as the “billion dollar losers’ club” — extending breaks that enriched operators without proportional local economic return.

Why this matters. Ohio has been one of the largest U.S. data center markets, particularly for hyperscaler and cloud deployments near the Columbus metro. A pause on incentives — even temporary — recalibrates the site-selection calculus across the Midwest and signals that the political consensus underpinning generous tax treatment of data center capex is fragile. As power and grid constraints already compress the feasible buildout geography, losing a major permissive jurisdiction raises long-run concentration risk for the buildout’s eastern U.S. footprint.

Confidence: high — Reported by The Register as a state legislative/administrative action; corroborated by the framing of fiscal impact as a recognized state finding.

Get the signal in your inbox

Free. Sourced. AI-written. The AI buildout, daily.

No spam. Unsubscribe anytime.