AI Revenue Counting Jolts Markets.

Share

The market's initial frenzy over AI revenue figures met a sobering reality check this week, forcing a critical re-evaluation of how growth is truly measured in this emergent sector. This fiscal clarity, or lack thereof, directly impacts valuations and the capital deployment strategies that will define the next phase of AI's expansion.

TL;DR

  • Market correction stabilises: Following Thursday's sharp decline, chip stocks recovered on Friday after Bloomberg indicated OpenAI's continued projection of a ~$70B net revenue run-rate by year-end, lifting Nvidia ~1.7% premarket and broader indexes.
  • Anthropic deepens API price competition: Anthropic introduced Claude Haiku 5.5, strategically cutting short-prompt API prices by up to 90% to match OpenAI's GPT-6 Luna rates at $0.10/$0.50 per million tokens.
  • Google’s unified enterprise agent: Google Cloud launched "Gemini agent," a single, autonomous assistant designed for persistent operation in the cloud, capable of spawning "coworker agents" with dedicated corporate email addresses.
  • OpenAI dismisses three safety researchers: OpenAI cites a "breach of trust" regarding sensitive data; the researchers contend their termination was due to prioritizing safety concerns.
  • Anthropic's "no cruelty" policy: Effective November 12, Anthropic's updated usage policy prohibits "sustained and needless" cruelty toward its models, alongside new restrictions on election, weapons, and surveillance applications.

Lead Story: AI Revenue Counting Jolts Markets

Thursday's market correction was not a technical capability failure, but rather a dispute over revenue accounting. By Friday, the broader market had largely rationalised the initial panic.

Reports indicated OpenAI's annualized revenue reached approximately $50B in September, falling short of the ~$70B figure previously circulated among investors (TechCrunch, Axios). The discrepancy was methodological: the higher number included gross partner revenue, aligning OpenAI's metrics with rivals like Anthropic, who account for sales booked via their cloud partners.

Markets reacted before clarification. CoreWeave dropped 7.8%, Oracle 5.5%, Broadcom 4.4%, and AMD 3.9%, with Nvidia alone shedding an estimated $170B in valuation (CNN). The trade reversed by Friday morning. Bloomberg confirmed OpenAI's internal projection of a net run-rate nearing $70B by year-end 2026, leading Nvidia to climb ~1.7% premarket to approximately $234.30, while the S&P 500, Dow, and Nasdaq each saw gains of ~0.4% at open (InvestingCube, GuruFocus).

The underlying signal remains: capital is increasingly flowing into private debt over public equity. Nvidia-backed Firmus reportedly cancelled a ~$30B Australian IPO due to insufficient demand (Bloomberg), while Broadcom is reportedly seeking over $50B in private credit to finance OpenAI's custom chip development (WSJ via InvestingLive). This timing is critical, as Anthropic prepares for its pre-IPO investor day on Tuesday, pursuing a valuation close to $2T. A week that exposed fundamental questions about AI revenue measurement forms an awkward prelude to the sector’s largest impending listing.

In Other News

Anthropic deepens API price competition. Anthropic unveiled Claude Haiku 5.5, strategically reducing short-prompt API pricing by up to 90%, now aligning with OpenAI's GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output (VentureBeat, Quartz). This aggressive discount targets prompts under 100K tokens, which account for approximately 90% of Haiku traffic, with longer prompts receiving a 50% reduction. Haiku 5.5 also introduces an adjustable effort setting and an expanded 1M-token context window, up from 200K.

Google ships a "universal agent for work." At Gemini at Work, Google Cloud launched Gemini agent: a single, autonomous assistant designed to answer queries, write and execute code, and operate across web, mobile, Workspace, Microsoft 365, and Slack (Google, SiliconANGLE). It maintains persistent operation in the cloud with shared memory and can provision "coworker" agents, each assigned its own @agents.company.com address and Workspace account. Financial and legal iterations are currently in preview, marking Google's most direct challenge yet to OpenAI's and Anthropic's enterprise footholds.

OpenAI dismisses three safety researchers; public dispute ensues. OpenAI confirmed the termination of Tomek Korbak, Jasmine Wang, and Mikita Balesni following an investigation that found a "significant breach of trust" concerning sensitive information, asserting the decision "was not about raising safety concerns" (CNBC, TechCrunch). The researchers dispute this narrative, claiming they were dismissed for prioritizing safety and cautioning against a chilling effect on internal dissent. This incident surfaces amid OpenAI's most intense period of scrutiny to date.

Anthropic bans "needless cruelty" toward Claude. An updated usage policy, effective November 12, explicitly prohibits "sustained and needless" abusive behavior directed at its models, while distinguishing ordinary frustration, critical feedback, dark fiction, and research (Forbes, The Register). Enforcement will remain light, with Claude terminating conversations as the primary mechanism. The revision also tightens restrictions on election interference, weapons, and surveillance applications.

X / Social Pulse

  • The "model welfare" debate goes mainstream as Anthropic’s cruelty ban attracts both endorsement and derision, with critics questioning the validity of "abusing" a statistical model (QZ).
  • Demis Hassabis adopted an optimistic stance in a recent documentary, positing AI's potential to "cure most diseases," a view contrasted with Tristan Harris's cautionary safety warnings (Democracy Now).
  • Jensen Huang, reflecting on success, stated, "Nobody suffers better than I can," adding that individuals with elevated expectations often exhibit reduced resilience (Fortune).
  • Yann LeCun reiterated his open thesis: training a frontier model is capital-intensive, while distilling one is low-cost, thereby compelling frontier models towards commoditization.

One to Watch

Anthropic’s pre-IPO investor day. This event on Tuesday, October 14, precedes a targeted mid-November listing aiming for a near $2T valuation. Observe how this week's revenue-counting controversy—particularly the methodology of booking partner-channel sales—influences investor inquiries. Also upcoming: Mistral’s Large 4 open weights, now confirmed for October 31.

Quick Hits

  • Pwn2Own Ireland: The event awarded $1,262,000 for 98 zero-days, notably including the first exploitation of a coding agent (OpenAI's Codex) alongside Oracle's Autonomous AI Database (ZDI, BleepingComputer).
  • Anthropic's Cyber Mission: The new initiative includes OSS Scanner, offering free frontier-model scans that identify bugs, generate working exploits, and propose patches without human review, supported by a $150M Genesis Mission pledge (unite.ai).
  • JetBrains open-sources Mellum 2.1: This 12B MoE "thinking" coding model (Apache 2.0) claims an 82% score on LiveCodeBench v6, outperforming Qwen's 75.4% (JetBrains).
  • Arena funding round: The LLM leaderboard secured $200M at a $3.1B valuation, with Lightspeed and Khosla leading the investment, signaling its expansion into AI-safety measurement (Bloomberg).
  • TSMC's record September revenue: The company reported NT$511.86B, representing a 54.6% year-over-year increase, with Q3 earnings scheduled for October 15 (CNBC).

The prevailing theme this week is less about AI's capabilities and more about the financial integrity of the ecosystem it supports. While Friday's rebound offered a measure of stability, a shelved mega-IPO, significant debt-funded buildouts, and an aggressive 90% reduction in inference pricing collectively shape the immediate backdrop for the sector's largest upcoming public listing.

Sources

Lock in. M. mazen@thorterminal.com

Read more