Perplexity revenue jumps 50% in pivot to AI agents- FT

Perplexity's revenues rose 50% in a month, driving estimated ARR to >$450M in March after launching a new AI agent and shifting to usage-based pricing. The start-up reports >100M monthly active users and tens of thousands of enterprise clients, monetizing via consumer and enterprise subscription tiers ($20–$200/mo). High-profile backers include Nvidia, SoftBank Vision Fund 2, NEA, IVP, Jeff Bezos and Yann LeCun, underscoring strong investor confidence in its agent-focused strategy.

Analysis

Perplexity’s shift to usage-based agent billing is a leading indicator for higher-frequency inference load that disproportionately values high-throughput datacenter GPUs and inference accelerators. In the next 3–9 months expect pronounced demand volatility: spikes driven by agent adoption will load factor cloud GPU fleets and create short windows of acute pricing power for suppliers of H100-class hardware and aftermarket spot capacity. That dynamic benefits makers of high-end accelerators and secondary markets that monetize spare GPU hours, while it raises the probability of customers exploring on-prem or edge-hosted inference to control variable costs.

Over a 12–24 month horizon, two offsetting trends will shape winners: (1) continued concentration of mission-critical, low-latency agents on premium silicon (positive for dominant accelerator vendors and ecosystem partners), and (2) faster-than-expected model compression, quantization, and optimized runtimes that materially reduce tokens-per-dollar for inference (negative for long-term gross compute growth). The inflection point where lower per-token cost meets broader agent adoption will determine whether cloud providers keep or cede margin to hardware vendors or enterprise customers.

Immediate tail risks that can reverse the trade are: a sudden increase in supply (new accelerator ramp or large cloud discounts) that collapses spot pricing within 1–3 quarters, or regulatory/enterprise governance constraints that cap agent automation and slow usage growth. Monitor GPU spot rates, datacenter lead times, and enterprise trial-to-paid conversion rates as high-signal, short-to-medium-term indicators of sustainability.

Key Decisions for Investors

• Long NVDA equity (6–12 months): add exposure on pullbacks into GPU-led weakness; asymmetric payoff if H100-class supply remains tight — target +15–30% upside vs downside limited by broad market; hedge with 5–7% portfolio allocation and stop-loss at 12% drawdown.
• Long NVDA call calendar (3–9 months): buy front-month calls and sell longer-dated calls to monetize theta while keeping upside if spot-driven demand spikes persist; position size sized to <2% notional to capture volatility re-pricing.
• Short high-multiple software/SaaS names that price infinite compute growth (3–12 months): rotate into names exposed to agent-hosting costs without hardware moats — use sector ETFs or single-name shorts to express conviction if spot GPU costs spike and margins compress.
• Tactical pairs trade (6 months): long NVDA / short AMZN (small size) to capture hardware vendor pricing power vs cloud pass-through risk — size 1:0.6 to reflect AWS revenue diversification; unwind if AWS announces aggressive custom-silicon discounts or major on-prem migration by enterprises.
• Risk control: set alerts for three catalysts to de-risk positions — (a) material increase in accelerator supply lead times shrinking below 8 weeks, (b) public cloud provider announcing sub-market pricing for inference instances, (c) enterprise agent churn or unexpected regulatory curbs. Close or hedge positions within 2–4 weeks of any trigger.