This week, Anthropic’s consumption-based AI revenue hit a staggering $47 billion annualized run-rate, eclipsing every public SaaS company except Microsoft, while Sail Research closed a $610 million round to slash AI inference costs, and Glean’s CISO warned of a 200% spike in AI-powered attacks.
This week, Anthropic’s consumption-based AI revenue hit a staggering $47 billion annualized run-rate, eclipsing every public SaaS company except Microsoft. The figure, confirmed by Financial Times on 30 June 2026, caps a meteoric rise from $9 billion in December 2025—a trajectory fueled entirely by charging per token, not per user. Meanwhile, Sail Research closed a $610 million Series A on 8 July to slash inference costs by up to 6x using queued, batch processing. And yesterday, Glean’s CISO hosted office hours to combat a 200% surge in AI-generated phishing and deepfake attacks reported on 5 July. Together, these events mark a turning point in how software value is created, delivered, and defended.
The Token Economy Arrives. On 23 June, venture capitalist Tom Tunguz analyzed Anthropic’s revenue data, revealing that the lab’s $47B run-rate already surpasses the combined trailing revenue of Palantir, Snowflake, and CrowdStrike. Anthropic never sold a single seat—instead, its Claude Code agent went from zero to a $2.5B nine-month run-rate by charging per developer’s agent workload. This consumption model decouples revenue from headcount, aligning cost directly with AI value. Enterprises are taking note: CIOs now ask, “How will you charge when my agent does the work?”
Async Inference Cuts Costs. Sail Research’s approach, detailed in a Theory Ventures blog on 7 July, leverages spare GPU capacity and fleet-aware orchestration to queue and batch-process tokens. For background tasks like code review and research, this yields dramatic savings versus real-time APIs. “Trillions of tokens are flowing through async systems,” noted Tunguz, predicting that by 2027, most inference will be queued. The $610M round signals investor conviction that the infrastructure layer must evolve beyond synchronous chat.
AI-Powered Threats Escalate. Dark Reading’s 5 July report of a 200% quarterly spike in AI-assisted attacks underscores the urgency that drove Glean’s 9 July office hours. Attackers now use frontier models for hyper-personalized phishing, deepfake audio, and automated reconnaissance. A mid-market firm lost $2M in June to a fake CEO voice, illustrating the need for AI-native defenses.
Seat-based SaaS pricing, pioneered by Salesforce in the early 2000s, revolutionized software by aligning cost with usage—but it assumed a human behind every screen. As agents multiply, that assumption crumbles. The shift to consumption pricing echoes the cloud’s transition from fixed servers to pay-per-use compute, but now it’s intelligence itself that’s metered. Similarly, cybersecurity has historically been reactive: signature-based tools catching known threats. The AI adversary era, however, mirrors the post-Stuxnet realization that advanced persistent threats outpace static rules.
These historical echoes clarify the stakes. Just as the cloud forced a security metamorphosis, the AI revolution demands a new defensive playbook—and a pricing model that scales with value delivered, not seats filled. Enterprises and startups that embrace token-first, async-first, and AI-resilient architectures will define the next decade. The evidence from the past month suggests the transformation is already underway.