Cognition vs Factory | Anthropic Under Threat | ElevenLabs Doubles Its Valuation to $22B
Summary
OpenAI’s reported Q3 reacceleration turns the frontier-model race from an Anthropic runaway into a fresh test of market expansion versus share rotation. Rory O’Driscoll contrasts roughly 70% quarter-on-quarter growth with just 18% from Q1 to Q2, while Dev Ittycheria says widespread switching driven by quality and price shows OpenAI is “back.” The “next shoe to drop” is Anthropic’s Q3 number: continued hypergrowth would validate a vast TAM, while a sharp slowdown would suggest OpenAI took share.
The Factory–Cognition dispute exposes a widening gap between legal freedom and reputational obligation in AI talent markets. Jason argues that “the definition of loyalty has changed” as executives and teams move faster, while Dev Ittycheria says access to confidential plans and a jump to a direct competitor can damage an executive’s standing long term. Vinod Khosla’s public description of Factory as a “struggling second-tier competitor” compounded the issue by handing rival venture firms “a weapon” against Khosla Ventures.
Reflection’s Beam could fill a strategically valuable gap: a US open-weight model near frontier quality at three to four times lower cost. Dev sees an enterprise proposition that lets customers “have their cake and eat it too,” particularly where Chinese models face regulatory or trust resistance. Jason thinks US models might capture half of open-source token flow within 12 months if they remain only three months behind frontier quality, though he cautions that published evals now require “three asterisks and four daggers.”
ElevenLabs at a $22 billion valuation earned Jason’s strongest investment call because voice quality remains less fungible than general reasoning. He would put up to 10% of the hypothetical fund into the company, citing widening product leadership, strong margins, large contracts and voice’s role as a critical agent interface. A flower-shop agent cannot merely be cheap: “It has to answer that phone in seconds, and it has to be the right flowers.”
Salesforce’s $2 billion purchase of Listen Labs validates AI-native market research, but not necessarily every standalone AI application. Rory sees a top-tier use case because LLMs can conduct adaptive voice interviews rather than march through canned surveys; Jason questions how roughly $20 million of revenue moves the needle at a company approaching $50–60 billion. Dev’s durability test is whether the startup has a proprietary data loop or is merely “a feature masquerading as a company.”
Agents are becoming a distribution channel with their own vendor preferences, making machine discoverability as important as human search rankings. Vercel reportedly gets 50% of new business from agents, up from 3% at the start of the year, while Resend’s MCP calls rose from 106,000 in April to 3.3 million in September. “When agents pick you, it’s a force of nature,” Jason says; the resulting opportunity for Profound, Peak and peers is to help companies shape the content and infrastructure that earn those recommendations.
The Groq–NVIDIA lawsuit could put license-and-hire transactions under legal and economic pressure if courts treat them as acquisitions in substance. Former engineers allege that technology and roughly 200 employees moved while some common shareholders retained equity in a hollowed-out company, challenging the principle that holders of the same security receive equal treatment. Jason’s formulation: “Just ’cause you say it’s not a duck, it don’t mean it’s not a duck.”
Aura’s pulled IPO is a warning that even rapid growth may not overcome pricing, secondary-sale optics and banker overpromising. Dev’s speculative but simple explanation is that demand did not support the promised valuation; Rory argues the board should perhaps have “hit the bid” once it had gone that far. With Anthropic’s IPO expected in mid-November, the broader lesson is that liquidity windows can deteriorate before an issuer finishes debating the perfect price.
Deep dive
1. OpenAI’s Q3 surge resets the frontier-model race
Dev’s framing is a comeback compressed into months: at the start of the year, many B2B and coding workloads appeared to work only on Anthropic; by the time of the discussion, OpenAI had products customers were adopting because they were “better and cheaper.” At approximately $70 billion of run-rate revenue, that speed of recovery is itself the signal.
Rory anchors the reversal in growth rates. OpenAI reportedly moved from 18% quarter-on-quarter growth between Q1 and Q2—which implied just under 100% year-on-year growth—to roughly 70% in Q3. At the prior checkpoint, the comparison was about $11 billion versus $6.8 billion, with Anthropic growing 10x against OpenAI’s roughly 2x.
The decisive question is whether OpenAI’s acceleration came at Anthropic’s expense. If Anthropic also posts strong Q3 growth, the result supports an enormous expanding TAM; if it falls from 10x toward 5x or lower, this begins to look like a rotating-share market. “OpenAI, you played your 70% card,” Rory says; now Anthropic must reveal its own.
OpenAI needed the acceleration because it has committed forward spending across compute, data centers and even land in Virginia. Rory calls the reported result “a big freaking sigh of relief”: without hyperscale revenue growth, many of those obligations could come unstuck.
2. Enterprise friction creates stickiness, but price and liquidity still rule
Dev separates developer behavior from enterprise adoption. Developers readily chase the “shiny new toy” or use several models at once; enterprises sign contracts, clear procurement, train thousands of employees and embed workflows. That makes replacement materially harder even when a superior model appears.
MongoDB’s internal conversations surface two enterprise constraints beyond benchmark performance: token budgets and IP rights. No organization has unlimited inference spend, while buyers must ask how models were trained, who receives their data and how proprietary information remains.
Against OpenAI’s proposed $30 billion raise at a $1.4 trillion valuation, Rory prefers liquidity at the margin. With the information set changing this quickly, he would rather own a public security than accept a private holding period that could stretch three or four years, regardless of an expected 2027 IPO.
Anthropic’s timing raises the stakes: the panel expects its Q3 figure to appear in the prospectus and be known before a possible mid-November listing. The growth disclosure—not merely the headline valuation—could decide whether investors see two compounding leaders or one taking share from the other.
3. Factory–Cognition turns implied loyalty into an investable risk
Harry lays out the allegation carefully: former Snowflake sales leader Chris Degnan was a Factory board observer, spoke to founder Matan daily and was close to him, according to Matan, while reportedly interviewing for Cognition’s CRO role. Dev’s conditional judgment depends on whether Degnan had access to confidential product, financial, board and competitive plans.
Jason’s counter is that perhaps 95% of CROs would view accepting a career-defining job as normal, particularly because Degnan was not a full-time employee or director. Rory adds that even full-time executives can join competitors carrying current knowledge; the adviser problem arose because expectations appear to have remained implicit rather than made concrete.
Dev rejects the idea that AI erases reputation. A leader has recruited people on a company, vision and financial opportunity; leaving for a direct competitor can make those employees wonder whether they were sold “a bill of goods.” Jason sees the old “rule of two” recruits disappearing as respected leaders now take eight or ten colleagues immediately, but Dev says human management remains much as it was in 1998.
Khosla’s intervention drew the clearest consensus. After leading Factory’s Series C and most recent round, Vinod Khosla publicly called it a “struggling second-tier competitor”; Dev says that gave every rival firm a deal-winning question: “Is this the partner you want when things go bad?” Rory’s verdict on competitive portfolio companies: investors must “love your children equally,” praise publicly and chastise privately.
4. Reflection’s opening is sovereignty plus good-enough economics
Dev’s case for Reflection’s Beam is specific: a US open-weight model near frontier intelligence, priced three to four times more cost-effectively, offers enterprises sovereignty, lower cost and less discomfort than Chinese models in regulated or data-sensitive industries. Jason’s Dreamforce conversations reinforce the demand signal: buyers using Chinese models often felt they were doing so “under duress for cost.”
Frontier parity is unnecessary for most workloads. If a model is within six months of the frontier yet covers 90% of coding, reasoning and agentic use cases, Dev sees little reason enterprises would not engage. The hard frontier model can remain reserved for “the hard stuff” while cheaper systems absorb routine volume.
Rory’s economic trigger is the absolute bill: if OpenAI and Anthropic each reach $70 billion of ARR, their combined $140 billion exceeds what Microsoft historically made across Windows and Office. At that scale, CFOs will order CIOs to shave millions from token spending, forcing routing across model tiers and open weights.
Jason could imagine US models taking 50% of open-source token flow within a year if they reach “X minus three months.” Rory says even 10–20% would be consequential because foundation models on OpenRadio receive perhaps 20–30% of the tokens but 90% of the dollars; enterprise-oriented US adoption would also de-risk inference platforms.
5. Falling token costs expand usage, while evals lose credibility
Dev does not see open weights and frontier labs as a zero-sum contest. Enterprise capability currently exceeds actual deployment because organizations lack experienced practitioners; as costs decline and teams gain fluency, Jevons paradox should bring more use cases, more consumption and room for multiple winners.
Demand is already outrunning current deployment capacity. Jason says every Dreamforce executive he met was overwhelmed by internal requests for tokens, with latent demand perhaps an order of magnitude above what companies have surfaced. Some spending will be wasteful or routed to suboptimal models, but the volume creates a substantial opening.
Safety will join cost and sovereignty as a procurement variable, though Dev concedes that defining it is difficult and “beauty is in the eye of the beholder.” Enterprises need some benchmark-like mechanism for trust, even if—as with performance tests—it can be gamed.
Jason’s warning is that Beam’s published results are not enough: “Every single eval needs like three asterisks and four daggers.” Speed, actual outputs and performance inside a mission-critical workflow matter more than narrowly optimized tests or “chess championships.”
6. ElevenLabs wins the rare 10%-of-the-fund recommendation
At a $22 billion valuation, Jason would allocate up to 10% of the hypothetical fund immediately rather than wait for another round. His case combines widening leadership despite intense competition, impressive margins and growth, “sub-millisecond response rates,” and voice becoming one of the ten most important underpinnings of agentic applications.
He has not abandoned his forecast that cheaper voice models will substitute for premium ones. His distinction is that ElevenLabs spans low- and high-end models while serving workflows where quality failure destroys the product: a phone agent that misunderstands an order or invents inventory has no value.
Reference calls showed even relatively modest companies happily spending hundreds of thousands of dollars annually. Jason therefore rejects blanket model fungibility: reasoning can be slightly weaker in a non-critical task, but real-time voice must work. “I don’t think anybody’s gonna catch them here today.”
7. Listen Labs makes market research an obvious AI-native category
Harry cites a $2 billion Salesforce acquisition, about $850 million returned on $96 million invested, a roughly 25x result for Sequoia across the seed and Series A, and a 4x return for Ribbit’s Series B in eight months. The speed and magnitude make timing as important as company-building duration.
Rory sees market research as a large, fragmented, pre-technical spend pool previously represented by consulting firms and companies such as Qualtrics, Medallia and SurveyMonkey. Voice is now a viable model interface, while LLMs can synthesize responses and alter subsequent questions dynamically.
His best contrast is the old survey that asks whether the bed was comfortable after a respondent says they never checked in. Listen Labs could replace that canned sequence with a conversational interview, making the application one of Rory’s “top five, top 10” AI use cases.
Jason understands the category but questions the buyer logic: roughly $20 million of revenue may not move Salesforce, which is approaching $50 billion and targets $60 billion. Dev expects more such purchases because distribution-rich platforms can absorb clever applications that are still features rather than independent franchises.
8. A $2 billion early exit is rational unless the company is generational
Dev’s durability test separates a copied feature from a franchise: users must create proprietary data, that data must improve the product, and the improved product must attract more usage. Competitors can copy functionality, but not the accumulated interactions powering that loop.
Harry contrasts Listen Labs’ three-year path with Intercom’s roughly 15 years of building. Rory refuses the envy frame—“life ain’t fair, and venture outcomes are very random”—and notes that timing could let 10% of today’s AI apps sell at 30x revenue while others grind toward six-times revenue with thousands of employees.
Jason would usually take $2 billion within five years unless the founders sincerely believe they are building “a Mongo or better.” Companies without a generational ambition have a terminal value, and a credible early offer can dominate the net present value of years more execution.
Dev’s BladeLogic history shows the choice recurring at every stage: Cisco’s John Chambers and EMC explored purchases before BMC ultimately paid about $900 million, roughly 9x the company’s $100 million of revenue, in 2008. He felt good continuing earlier and equally good selling when the eventual moment arrived.
9. Agents are becoming the buyers that determine infrastructure winners
Vercel reportedly reached $600 million in ARR while agents rose from 3% to 50% of new business. Jason recalls roughly 170% growth alongside something like a 20% increase in customers, implying extraordinary expansion as autonomous systems spin up applications and consume infrastructure.
The advantage was not designed in hindsight: Vercel predates the agent wave, just as Replit dates to 2016. Both had the right substrate when demand shifted. “When agents pick you, it’s a force of nature,” Jason says, because their preferred stack is surprisingly difficult for human operators to override.
Resend is his concrete example. After SendGrid setup repeatedly failed, Jason asked an agent what to use and received Resend; separately, Resend’s MCP calls rose from 106,000 in April to 3.3 million in September. That is not conventional lead generation—it is software directly recommending and activating another product.
An agent’s choice is also more contextual than a generic ChatGPT answer. It knows the application, infrastructure and dependencies, then selects for that stack; Jason’s own agent recommended Render over Railway, Fly.io and Vercel for one project, creating a qualified lead without the buyer previously knowing Render.
10. Machine visibility could become a deeper market than SEO analytics
Dev recasts 25 years of search optimization: companies once fought to appear on Google’s first page; now they must ensure agents can find them, understand their product and access current, accurate information. “If you’re missing from an AI-generated answer, it’s like being invisible on Google.”
Startups face an asymmetric corpus problem. Incumbents have 10–15 years of documentation and information, while challengers may be nearly absent from model context. To become a first-class agent choice, a startup must expose excellent documentation, APIs, integrations and infrastructure.
Profound and Peak can become more than monitoring dashboards if they close the loop: diagnose where a vendor is absent, then shape and create the content that fixes the absence. Rory compares the expansion path to HubSpot, which began with an immediate SEO pain point and layered adjacent marketing problems into a platform.
Jason proposes a simple field test before the full product exists: ask ten trusted teams with production agents, every two weeks, what those agents recommend in a category. The emerging market is not merely tracking generic LLM responses; it is observing stack-aware decisions made by live agents doing real work.
11. The Groq case challenges license-and-hire deals at their foundation
Harry introduces a $17 billion Groq–NVIDIA license-and-hire dispute; the complaint, as described, alleges an $11 billion license plus roughly $3 billion in stock, with core technology and around 200 employees transferred while common holders retained a hollowed-out company. Two former engineers brought the Delaware action.
Rory’s legal intuition begins with equal treatment: common shareholders own the same security, and directors owe duties to all of them. A buyer cannot ordinarily allocate rich consideration to employees it wants while ignoring former employees holding identical shares. That makes the plaintiffs’ “clever little case” potentially settlement-worthy.
Dev expects employees to begin asking a question investors already negotiate: what happens to my equity if the technology and selected staff are bought but the corporate shell is not? Rory warns that once established common-law protections erode, every employee must price increasingly elaborate transaction scenarios before joining.
The structure exists because a conventional acquisition might face nine months of regulatory review while a license-and-hire can close on Sunday and move employees on Monday. Jason thinks meaningful legal liability, added to double taxation, could end the practice: “Just ’cause you say it’s not a duck, it don’t mean it’s not a duck.”
12. Shipping separates Muse from Dots, while Aura tests the IPO window
Harry praises Muse’s launch and calls Dots’ enterprise-travel demo tech-centric and underwhelming. Rory, previously skeptical of Meta’s execution, credits the team for resourcing Muse with compute and shipping a state-of-the-art product: “They put their heads down and shipped product,” which “cures everything.”
Jason asks for a developer rather than consumer verdict on Dots. Its initial value may be integration with Codex as a persistent 24-hour coding agent. He notes that Replit and Lovable could stay on in the early days, while Claude Code and Codex are hard to keep running 24 hours; he says that limitation helped OpenClaude take off. He would allow 30–90 days before judging it by a weak travel-booking presentation.
Aura supplies the capital-markets counterpoint. Jason is troubled that a company described as doing $1.2 billion, growing 74% with an attractive recurring-revenue portion, pulled its IPO; Dev, explicitly speculating, thinks pricing failed, while Forerunner’s full exit hurt optics. Rory’s prescription is blunt: distrust banker optimism, “hit the bid,” and remember how much can change before Anthropic’s expected mid-November offering.