Chasing IPO Gains
in AI is a
High-Wire Act
Anthropic's pending IPO, China's open-weight price war, a $725 billion annual infrastructure bet, and a bifurcated economy — stress-tested for institutional portfolios.
An 80-fold revenue surge meets a 30x valuation ask
Anthropic's annualised revenue ran at roughly $1 billion at the end of 2024. By December 2025 it reached $9 billion. By April 2026, the company confirmed $30 billion-plus in ARR — an 80-fold increase in roughly 16 months.
The Series G closed at a $380 billion valuation. Anthropic is now raising $40–50 billion targeting a $900 billion valuation. Secondary and on-chain markets imply $1.0–1.2 trillion. At $30B ARR and a $900B valuation, the implied revenue multiple is 30x.
One structural wildcard sits behind all of this: S&P Dow Jones closed a public comment period on May 28 for a rule change that would allow mega-cap companies to join the S&P 500 in six months rather than twelve.
One confirmed data point.
Everything else is inference.
The Series G at 12.7x ARR is the only closed, reported transaction. The $380 billion valuation is audited. Everything above it — the $900 billion fundraise target, the $1.0–1.2 trillion secondary-market price — is either a stated target or a market-implied figure derived from illiquid secondary trades.
IPO bankers have floated a $400–500 billion range, implying 13–17x ARR. That range would represent a modest premium to the Series G and a significant discount to the secondary market. The divergence is the crux of the valuation debate: the secondary market is pricing in sustained hyper-growth and margin expansion to 77%; the bankers are pricing in a more modest trajectory with open-weight competition as a ceiling.
For context, Snowflake peaked at roughly ~40x NTM revenue in 2021 — during a zero-rate environment with no credible open-source alternative to its product. Anthropic faces open-weight competitors today that did not exist for Snowflake then.
China trained the competitor
for $5.6 million
DeepSeek trained V3 for a reported $5.6 million in compute costs — versus $50–100 million estimated for GPT-4. The model matches frontier performance on most benchmarks at a fraction of the training cost. That gap is the supply-side shock that benchmarks miss.
The demand-side implication is direct. Claude Opus 4.7 output costs $25 per million tokens on the Anthropic API. DeepSeek V4 costs $3.48 per million — an 86% cost reduction available today without self-hosting. LLaMA 3.1 70B self-hosted runs near $0.08 per million tokens — a 312x gap versus Claude.
Use the toggle above the chart to see the full enterprise cost comparison at 100 million tokens per month — the scale at which the pricing gap becomes a board-level procurement decision rather than an engineering preference.
The largest private infrastructure build
in U.S. economic history
Amazon, Microsoft, Google, and Meta combined plan $725 billion in capital expenditure in 2026 — up 77% from $410 billion in 2025. Roughly 75% of that, or approximately $544 billion, is directed at AI infrastructure: data centers, networking, custom silicon, and model training capacity.
No single peacetime infrastructure programme — interstate highways, rural electrification, broadband build-out — has deployed private capital at this rate. The hyperscaler bet is not a hedge. It is a committed position on AI becoming the dominant compute workload within three years.
Anthropic sits at the center of this spend in two ways. It consumes hyperscaler compute as a customer. And it has committed $80 billion in cloud fees to Amazon, Google, and Microsoft through 2029 — a fixed-cost obligation that becomes a gross margin drag if per-token pricing compresses faster than revenue scales.
The $725B capex cycle is also a systemic concentration risk. If AI revenue disappoints — through pricing erosion, utilisation shortfalls, or open-weight substitution — hyperscalers face simultaneous earnings misses and asset impairments. The GDP contribution from AI-adjacent construction and equipment investment then reverses, exposing the bifurcated economy beneath. This is an Icarus Asia estimate.
Three paths for AI unit economics
The valuation question ultimately reduces to a single empirical variable: at what rate does per-token pricing compress, and what gross margin does Anthropic achieve at each price level? The charts above show three scenarios. Each carries materially different portfolio implications.
The base case — 50–60% probability — assumes Anthropic's enterprise moat holds. Safety regulations, compliance requirements, and switching costs keep the 80% enterprise revenue share intact. Pricing falls 15–30% but gross margin expands toward 40–55% on volume. IPO at the banker range of $400–500 billion is achievable.
The moderate scenario — 30–40% probability — assumes open-weight models capture the commodity API tier faster than expected. Per-token pricing falls 40–60%. Gross margin stalls at 35–45%. IPO is delayed or repriced downward. The $80B cloud commitment becomes an anchor rather than an asset.
Click any scenario card in the panel to expand the portfolio implication for each path.
Three questions for every
investment committee
Before any institutional participation in Anthropic's IPO or secondary market, three analytical questions must be answered with primary data — not market consensus.
One. What is the revenue split between safety-regulated enterprise contracts and commodity API access? Apply a 15–20x multiple to enterprise revenue separately from a 5–8x multiple on commodity API. The blended multiple at the current mix implies a materially lower fair value than the $900B headline.
Two. What is the net margin impact of the $80 billion cloud commitment through 2029 under each pricing scenario? If per-token revenue compresses 50%, does the committed cloud spend become a fixed loss-making obligation?
Three. At what level of open-weight adoption does gross margin plateau below the 77% target — and what is the IPO multiple at that plateau? A 40% gross margin at $30B ARR implies a very different terminal value than 77%.
Do not chase the secondary market price. The $900B target implies the base case with full credit for the 77% gross margin trajectory — and zero allowance for open-weight disruption, cloud commitment drag, or macro softening. The Series G at $380B (12.7x ARR) reflected a rational risk-adjusted entry. The current secondary price does not. Wait for the IPO prospectus and apply a sum-of-the-parts multiple to the enterprise and commodity revenue segments separately. At any price above $500B, the risk-reward requires explicit commitment to the base case and documented tolerance for a 40–60% drawdown in the moderate scenario. This is an Icarus Asia estimate · Not investment advice.