The Investor

The contradiction in the coverage

The same news cycle tells two opposite stories about DeepSeek. The Information's dealmaking coverage still casts it as the price-pressure player: open-source models and price cuts are keeping AI costs in check, and DeepSeek's low-cost scale adds to that pressure. The Information's own reporting on the raise says something different. Revenue more than doubled partly because DeepSeek raised prices. Both can't be the right lens for your cost models. The pricing behavior is the harder evidence, because a company can inflate a revenue number far more easily than it can fake demand holding up after a price increase.

Treat the revenue figure with care. It is an unaudited, CEO-supplied annualized run-rate, disclosed at a private investor meeting and leaked during an active raise. The size of the price hike was not disclosed. Rely on the direction of the move, not the magnitude.

Who takes the tier DeepSeek left

A price increase from the cost leader leaves room underneath it. AINews data shows who is moving into that room. Xiaomi's MIT-licensed MiMo-V2.6-Pro scores 46 on the AA index against GPT-5.6 Sol's 47, at $0.13 versus $1.99 per task. In practice, the ultra-low-cost position is moving from one Chinese lab to an open-weight release from another. For sourcing, the low-cost tier is not closing. It is being taken over by open weights, which favors serving-efficiency and routing companies over any single model vendor.

Why the 75x doesn't transfer

The deal terms reported by The Information are 50B yuan ($7.5B) at 500B yuan (~$75B), roughly 10% dilution, a targeted end-October close, and a Shanghai Stock Exchange listing path. The Information Briefing notes DeepSeek is raising cash equal to 7.5x its revenue run-rate. Three features make the multiple specific to DeepSeek:

• Captive capital. The buyers are domestic, and the listing venue has no Western institutional access.
• A restricted TAM. Choosing Shanghai confirms continued exclusion from US and EU regulated enterprise buyers.
• A thin reprice cushion. At ~10% dilution, there is little room to adjust if the close slips.

Founders will still bring this number to your next negotiation. The committee note should come first.

The comp worth importing from DeepSeek is its pricing behavior, not its valuation multiple.

The disclosure event ahead

A Shanghai listing would make DeepSeek plausibly the first frontier lab with audited public financials. The Information notes that Anthropic's IPO timing leaves Wall Street without a Western comparable. Audited gross margins and capex intensity will either support private AI marks across the sector or put them under pressure. Your markdown policy should be agreed before that happens.

The smart move

Model AI monetization and AI cost deflation as two separate variables. Most sector models still assume they move together. Then stress-test the holdings whose path to software-like gross margins depends on per-token costs falling.

Money no longer buys capacity

The most important line in The Pragmatic Engineer's reporting comes from a VP of Engineering at a large inference provider. The company has cash and is willing to sign the longest leases on offer, and its cloud providers still say there is no capacity. Allocation, not capital, is the binding constraint. That changes what a funding round buys. A portfolio company that raises to scale agent workloads may find the money can't be turned into compute on its planned schedule.

Two named operators support the claims, which come from a dinner with CTOs and heads of infrastructure. turbopuffer CEO Simon Eskildsen says "getting CPUs is not easy anymore." Katelyn Lesse, who leads platform engineering for Anthropic's Claude Platform, puts server prices up 10-20% and fulfillment at about 6 months. Customers are prepaying for capacity arriving in December.

The mechanism: every agent step runs on a CPU

Agents compile, test and lint code. RL training runs models inside software environments. Uber's agentic requests grew 9x in six months, and Uber and Ramp have moved coding agents off laptops and onto dedicated cloud instances. That turns compute employees already had into recurring metered spend.

AINews data shows the same pattern from the token side. Anthropic cut Opus 5.5 list prices to $4/$20 per million tokens, yet cost per coding task rose to $13.04 because the model uses 15.6M tokens per task. Per-unit prices are falling while per-task bills rise, and the CPU tail comes on top of that.

The arithmetic on spot capacity is severe. A job bought at a 90% discount was paying 10% of list, so moving it to list price is a 10x cost increase on that line.

How long it lasts, and what to discount

Supply is squeezed at two points. TSMC splits its production lines across GPUs, CPUs, Apple, Qualcomm and Broadcom. Memory makers are shifting wafers to HBM, which makes the DRAM every CPU server needs more expensive. Analysts cited by Lesse expect CPU relief before memory relief, but still multiple quarters out, which likely means into 2027. Discount appropriately: the evidence is high-quality anecdote, the claim that some regions are closed to new tenants is explicitly a rumor, and paid subscribers received the report about 14 days earlier.

When allocation replaces capital as the constraint, a funding round buys less growth than the model says it does.

The smart move

Favor software that pays off whether CPUs are scarce or plentiful: capacity forecasting, workload consolidation, sandbox density, and reserved-capacity secondary markets. Be wary of CPU resellers underwritten on scarcity pricing. They face the same TSMC and DRAM limits, and CPU supply is expected to normalize first.

What actually moved the price

The repricing followed a conference moment, not a financial disclosure. At AI Agenda Live on Sept. 23, Nvidia's Dion Harris called Jev "incredibly fast" while noting its limitations. Atlassian's Tamar Yehoshua called it "so much cheaper and so much faster" but less accurate than leading LLMs. The Information notes the talks are early, the sources are unnamed, and TypeSafe declined to comment. Treat $10B as an offer, not a print.

Jev outputs numbers and calibrated confidence scores instead of prose. That makes it a model for machine-to-machine jobs such as classification, routing and agent monitoring. The category is real. The question is whether one vendor can hold pricing power in it.

Three sources, three different cost ratios

The ratio shrinks as the comparison gets cheaper. AINews also reports that a classic BGE-small plus logistic regression baseline beats Jev on Banking77 (93.3% vs 83.2%), and that the open-weight AutoJev-27B sits 0.8 points behind it on a decision index. TLDR Data adds that a separate audit found benchmark defects skew toward inflating scores, which makes interested-party benchmarks especially weak collateral.

The underwriting math

The Dealmaker analysis assumes a 25x revenue multiple. On that basis, $10B needs about $400M in revenue. Earning that only by displacing spend at 1/400th pricing would mean absorbing roughly $160B of frontier-equivalent workload, likely more than the entire frontier-API market. AINews reaches the same conclusion from list pricing. At $0.044 per 1K judgments, even $100M of ARR takes about 6.2B judgments a day. Cascades also hand most of the savings to buyers: a Jev-to-frontier cascade keeps 99% of accuracy at 57% of cost, and nearly all of that 57% still goes to the frontier model.

A $10B price on a decision model is a bet that agents create demand for machine judgment that doesn't exist yet.

The bull case rests on one line from Yehoshua: Atlassian sees uses "for places we're not using an LLM because it wasn't cost-effective." That is a Jevons bet, meaning cheaper judgment could expand total consumption. It is plausible, but no revenue has shown it yet.

Where the upside actually was

The winners in this story entered at $200M. Core Automation shows the pricing logic. It raised $630M at a $3.5B post-money with an elite ex-OpenAI founder, yet it would be priced below TypeSafe, because practitioner proof beats pedigree. The likely exit route is strategic, with compute as the currency. Tworek said he would consider a deal with anyone offering "all the GPUs in the world."