In just 30 days, token consumption overtook per-seat software, batch inference slashed AI costs, and AI attacks doubled—mandating a unified enterprise response.

This week, Sail Research landed $610M to slash AI costs 6x with queued inference, signaling a tectonic shift in how enterprises run background agents—and alongside Anthropic’s $47B run-rate and surging AI attacks, the old playbook is obsolete.

July 10, 2026 — The tectonic plates of enterprise technology shifted violently in the past 30 days, and the tremors are not yet done. When Sail Research disclosed a $610 million Series A on July 7, it wasn’t just a funding round; it was a loud declaration that the future of AI inference belongs to patient, batch-processed background agents—not the real-time chat interfaces that have dominated headlines. Combined with Anthropic’s breathtaking $47 billion revenue run-rate—powered entirely by token consumption—and a doubling of AI-driven cyberattacks in Q2, the message is unmistakable: companies must now navigate a token-to-threat continuum, where pricing models and security postures converge.

Current Waves (since June 10, 2026)

The exponential trajectory of Anthropic has turned software economics on its head. As Tom Tunguz noted on June 23, by year-end Anthropic will out-earn every public software company except Microsoft. In December 2025, the company’s run-rate was a mere $9 billion; by mid-May, it hit $47 billion. Unlike Salesforce, Adobe, or ServiceNow, Anthropic never sold a seat. It sells tokens—the raw fuel of AI agents. Claude Code alone went from zero to a $2.5 billion nine-month run-rate by charging per developer’s agent workload. This decouples revenue from headcount, aligning cost directly with value delivered. For CIOs, the question is no longer “How many seats do I need?” but “How do you charge when my agent does the work?”

Just as the pricing floor crumbled, the inference cost ceiling got a sledgehammer. Sail Research’s model—using commodity GPU spot capacity, queuing requests, and fleet-aware orchestration—promises up to 6x cost reduction versus real-time APIs. As founders and investors placed their bets on July 7, the implication was clear: the majority of future tokens will flow through asynchronous, batch-processed systems. Background tasks like code review, data enrichment, and deep research don’t need sub-second latency; they need throughput and affordability. Sailboxes keep agent state alive for persistent AI workers, charging only for active compute. This bifurcates the inference market into real-time and batch tiers, each with distinct infrastructure demands.

Yet while innovation races forward, so do adversaries. On July 9, Glean’s CISO Office Hours spotlighted a menace that has doubled in force: AI-powered cyberattacks. According to Dark Reading’s July 5 report, AI-generated phishing campaigns surged 200% from Q1 to Q2 2026. Attackers leverage frontier models for hyper-personalized reconnaissance, flawless lures, and deepfake voice calls that can trick even trained employees. Last month, a mid-market firm lost $2 million when a CFO’s cloned voice authorized a wire transfer. Traditional rule-based filters and annual training are no match for an enemy that automates context-aware deception.

Historical Echoes

This trifecta of disruption is not without precedent. The cloud revolution of the 2010s similarly upended software pricing—from perpetual licenses to SaaS subscriptions—and forced a security paradigm shift. When Salesforce introduced its pay-per-user model, incumbents struggled to adapt. Today, token-based AI replicates that disruptive arc but at hyper-speed. The difference: consumption pricing scales with machine output, not human headcount, creating an entirely new revenue geometry. Just as cloud adoption necessitated the shared responsibility model and CASB tools, the ascent of agentic AI demands a reimagined defense stack—one where zero-trust isn’t a buzzword but an architecture for an era when any interaction could be synthetic.

Historical parallels in cybersecurity are equally stark. The rise of ransomware-as-a-service in the early 2020s democratized extortion; now, generative AI democratizes deepfake-powered social engineering. The same way companies scrambled to deploy endpoint detection and response (EDR) after seeing waves of cryptolockers, today’s defenders must adopt real-time AI detection tools that analyze intent and context, not just signatures. The Glean Office Hours echo an urgent industry cry reminiscent of the 2017 WannaCry aftermath—only faster and more insidious.

The convergence of these three forces—token-based pricing, asynchronous intelligence, and AI-native threats—forms a continuum. An enterprise that adopts Anthropic’s API to power a code assistant must architect its infrastructure for batch inference to stay cost-effective, while simultaneously hardening defenses against the very language models it’s using being weaponized against it. This is the new normal. Founders must redesign pricing around value delivery, not seats; CTOs must embrace queued inference for massive agent workloads; and CISOs must embed AI-aware forensics into their SOCs. The window to act is measured in weeks, not years. As June 2026 bleeds into July, the companies that will thrive are those that grasp that the token driving the agent is the same token the attacker will exploit—and that the only viable response is a holistic, integrated strategy spanning commerce and security.