Fable 5.2 cuts prices by 20%, Jensen Huang resigns, and Nebius raises new-deal GPU-hour prices by over 20%
Author
AgentsFlare Research
Date Published
Model prices and compute procurement terms moved in different directions this week. Anthropic reduced Fable 5.2's standard input and output prices by 20%, while Google's Gemini 4 Hydrogen entered the everyday agent market at $1.50 per million input tokens and $7.50 per million output tokens. Further upstream, Nebius reported that average GPU-hour prices on third-quarter new deals rose over 20% sequentially. Prepayments on new contracts lasting at least a year covered more than 100% of associated capital expenditure. CoreWeave expects approximately $9 billion in customer prepayments and deposits over the next three months. In the same week, Anthropic launched its IPO roadshow, and NVIDIA announced Huang's resignation, with CFO Colette Kress also serving as interim CEO. API buyers can still benefit from lower usage-based prices. Buyers reserving dedicated capacity may have to fund it before taking delivery. A supplier's ability to raise capital and deploy equipment on schedule is becoming another procurement consideration alongside price.
If you read nothing else this week:
- NVIDIA changes its chief executive. Jensen Huang resigns as president, CEO and director. CFO Colette Kress also becomes interim CEO, and the board begins a successor search. (Announced November 12)
- Google expands its eighth-generation TPU systems. TPU 8tx targets confidential inference. TPU 8i Max combines two TPU 8i chips in one package for large-context workloads. (November 9)
- Fable 5.2 cuts standard API prices by 20%. Input costs $8 per million tokens and output $40, versus $10 and $50 for its predecessor. (November 9)
- Gemini 4 Hydrogen targets everyday agent workloads. Standard pricing is $1.50 per million input tokens and $7.50 per million output tokens across the full one-million-token context window. (November 12)
- Alibaba lowers subscription costs and opens a limited Qwen4-Max preview. Token Plan Personal Edition Pro's discounted website price falls to RMB312 per month, with monthly Credits rising from 180,000 to 200,000. Qwen4-Max-Preview runs through November 25 at 24:00 Beijing time. (November 9 and 11)
- Decision models attract production traffic. Vercel reports that they handled 26% of its gateway's October requests while accounting for 0.53% of total spending. (Published November 9)
- Anthropic launches its IPO roadshow. The preliminary range is $70–78 per share for 825 million shares offered by the company and selling stockholders, alongside a conditional $26 billion private placement. (November 10)
- CoreWeave grows revenue, with net interest expense still above operating income. Third-quarter revenue is $3.661 billion, operating income $207 million and net interest expense $846 million. (November 10)
- Nebius raises new-deal prices; Nscale discloses further financing needs. Nebius's third-quarter new-deal GPU-hour prices rise over 20% sequentially. Nscale's amended prospectus says it needs additional financing to meet forecast obligations in 2027. (November 12)
- Microsoft 365's disruption extends into a second day. Affected services include Outlook, Teams and Microsoft 365 Copilot. Microsoft says services are fully restored on November 14. (November 12–14)
AgentsFlare is the enterprise AI control plane. As models, clouds and agents fragment, it keeps routing, cost attribution and access auditing under enterprise control.
AI Infra Weekly is AgentsFlare's strategic column for enterprise users, tracking consequential shifts in global AI infrastructure. A control plane has no inherent allegiance to any single model or cloud. That independence lets us assess structural changes in models, compute and regulation without taking sides—and identify their direction before the landscape becomes fixed.
Huang resigns; multicloud procurement can still depend on the same upstream supplier
On November 12, NVIDIA announced Huang's resignation and appointed CFO Colette Kress interim CEO while retaining her finance responsibilities. The board will consider internal and external candidates for a permanent successor. The announcement explicitly said operations, customer commitments and the product roadmap were expected to continue without interruption. The leadership change does not establish that chip deliveries will be delayed, nor does it invalidate existing orders.
NVIDIA's influence now extends beyond chip delivery. Anthropic's offering documents this week list a proposed $10 billion private-placement investment from NVIDIA; Nscale's filing also describes NVIDIA financing. When the chip supplier participates in customers' capital arrangements, procurement and financing relationships overlap. An enterprise reserving capacity from several GPU clouds may still depend on the same chip company and its financial support. More contracts do not automatically reduce that shared exposure.
Another supply development came from Google. TPU 8tx and TPU 8i Max, announced November 9, are additional configurations within the eighth-generation TPU portfolio, targeting confidential inference and large-context workloads respectively. Google says its TPU business now has nearly 100 customers. These systems add hardware options for some workloads, but migration still depends on software compatibility, regional availability and delivery terms. Assessing supplier independence requires looking beyond the calling endpoint to the cloud, hardware and capacity actually serving the model.
Fable 5.2 remains twice Opus's price after its cut; Hydrogen pursues everyday calls
Anthropic released Fable 5.2 on November 9. Standard API input costs $8 per million tokens, output $40 and cache reads $0.20. Fast mode costs $16 for input and $80 for output; Anthropic says it can run up to 2.5 times faster than standard mode. US-only inference adds a 1.1 multiplier to input and output prices. The 20% reduction compares predecessor standard prices; it does not describe every operating mode and deployment scope.
The capability gains also have task-specific limits. In Anthropic's published evaluations, FrontierSWE v2 rises from 56.3% for the predecessor to 69.9%, but Fable's AA-Briefcase v1.1 Elo is 1785, below Opus 5.5's 1800. These are vendor-published results, not independent validation across all enterprise workloads. Anthropic itself still recommends starting with Opus 5.5 for most everyday coding and knowledge work. Opus's standard input and output prices are $4 and $20, half Fable's. Moving everyday workloads wholesale to the newest flagship can therefore increase the bill even as that flagship becomes cheaper than its own predecessor.
Google released Gemini 4 Hydrogen on November 12 at $1.50 per million input tokens and $7.50 per million output tokens across its full one-million-token context window. Both rates are 18.75% of Fable's standard prices. Google positions Hydrogen for everyday coding and agent execution, with Carbon and Argon retaining deep, long-horizon reasoning roles. Both vendors are segmenting tasks and prices. The rate difference alone cannot establish which is cheaper at equal quality. It does, however, change the economic basis for sending every task to an expensive flagship: difficult judgment, routine execution and simple decisions increasingly have distinct price points.
Qwen4-Max enters through a subscription preview, with quotas and an expiry date
On November 9, Alibaba Cloud raised quotas across all four Token Plan Personal Edition tiers. On its China website, Pro's discounted monthly price fell to RMB312, while Credits increased from 180,000 to 200,000. Qwen4-Max-Preview joined Personal and Team editions on November 11, with the preview ending November 25 at 24:00 Beijing time. Promotional Credit consumption is as low as 20% of the standard rate, with a further 60% discount between 22:00 and 08:00, taking the minimum to 8%. These are subscription-credit rules, not directly comparable market-wide API prices per million tokens.
The Team Edition documentation also says the model will be retired after the preview. Enterprises receive a two-week evaluation window; long-term availability remains subject to subsequent arrangements. The promotion lowers experimentation costs. For a business directly dependent on the preview model, the deadline also creates a required migration.
Subscriptions change how costs are assigned. Token Plan pools multiple models under Credits, assigns Team Edition seats to members and offers shared usage packs. Switching within the platform becomes easier, but enterprises must still consolidate that spending with other clouds, direct APIs and private deployments. A supplier's unified bill has a boundary. When the same platform controls the model catalog, credit discounts and availability period, independent cost attribution and call records give the enterprise a basis for deciding whether renewal remains worthwhile.
Decision models handle a quarter of requests, giving routing measurable economic scale
Vercel's production index, published November 9, reports that decision models handled 26% of its gateway's October requests but accounted for only 0.53% of total spending. Nearly a third of paid teams used one; adoption reached seven in ten among teams processing more than 10 billion tokens per month. The gateway offered 15 decision models from eight vendors. This is one platform's anonymized aggregate sample, not a measure of the entire enterprise AI market.
The data adds adoption evidence to the product announcements. Classification, scoring and tool selection are separating from long-form generation. Low token prices and very short outputs create a substantial gap between request share and spending share. Evaluating AI use only through token totals or the overall bill can understate the influence of these decision steps. They consume little while determining which expensive model receives a request or which tool is called next.
This is why AgentsFlare focuses on who controls the decision. Decision models can come from any provider; routing rules, project budgets and call auditing should remain under unified enterprise management. The platform supports spending caps by key and project, real-time usage tracking and immutable audit logs. Enterprises need to see both the call fees saved and the destination of the traffic to establish whether a cheap decision creates more expensive downstream work.
Anthropic begins its roadshow, with Google and NVIDIA proposing $10 billion each
On November 10, Anthropic announced its IPO roadshow at a preliminary $70–78 per share. Of the 825 million shares offered, approximately 801 million are newly issued by the company; the rest are sold by existing stockholders, whose proceeds do not go to Anthropic. The prospectus separately lists a $26 billion concurrent private placement from seven industry partners, including $10 billion each from Google and NVIDIA. The placement depends on completion of the public offering and other conditions. At the $74 midpoint, excluding the underwriters' additional-share option, Anthropic estimates combined company net proceeds of approximately $84.6 billion.
Google's purchase exercises an option under an existing investment agreement. The full placement should therefore not be presented as an investment commitment that suddenly arose this week. These also remain proposed offering terms: the final price and proceeds have not been determined.
The source of capital matters to enterprise buyers. Investment from cloud and chip partners can fund a model company's infrastructure expansion while further intertwining supply and capital relationships. Multiple model, cloud and chip brands can share investors and long-term procurement commitments behind the scenes. Completed financing can strengthen resources available for delivery; it does not demonstrate that every cloud contract has become usable capacity. Capital receipt, equipment deployment and service availability need separate assessment.
CoreWeave's net interest expense is four times operating income; Nebius funds construction through prepayments
CoreWeave reported third-quarter revenue of $3.661 billion on November 10, up 168% year on year, with operating income of $207 million. Net interest expense was $846 million, approximately 4.1 times operating income, and the quarterly net loss was $709 million. Adjusted EBITDA margin reached 62%, but that measure excludes interest, depreciation, amortization and other items. A capital-intensive compute provider can report a high EBITDA margin and a bottom-line loss at the same time. Equipment investment and financing costs still affect returns.
Management expects approximately $9 billion in customer prepayments and deposits over the next three months. More than $3 billion relates to recontracting existing infrastructure at prices above the original contracts. Existing equipment therefore retains pricing power at renewal. Advance cash also carries future service obligations and cannot be counted as current revenue or profit. Earlier receipt of customer funds supports supplier expansion while increasing the interval between buyers' payment and actual consumption.
Nebius disclosed more direct evidence on November 12. Average GPU-hour prices for third-quarter new deals rose over 20% sequentially. More than 90% of signed contract value carried a prepayment. For new agreements lasting at least a year, prepayments covered more than 100% of associated capital expenditure, versus 50%–60% in the second quarter. These figures describe the relevant Nebius contracts; they are not a market-wide GPU price index.
Better collections did not remove delivery constraints. Nebius said some third-party capacity had slipped into the first half of 2027, leaving expected 2026 revenue at the low end of its existing $3.0–3.4 billion range. Together, the disclosures show that strong compute demand and larger prepayments can coexist with delivery limits. Site readiness and equipment deployment still determine when revenue materializes. For the buyer, the same process determines when advance payments become usable resources. The time between payment and use, capacity acceptance, and remedies for delay all carry real costs alongside the hourly rate.
Special insight: Nscale increases equipment purchases, with 2027 funding still to be secured
UK AI infrastructure company Nscale updated its listing documents on November 12, estimating nine-month revenue of $307.5 million, an operating loss of $837.5 million and a net loss of $1.6919 billion. These are preliminary, unaudited figures. The net loss includes $629.9 million in fair-value adjustment losses and should not be treated entirely as an operating cash outflow.
The filing says Nscale increased GPU, networking and related equipment purchases to mitigate price and allocation risks and the delay penalties specified in its Anthropic service agreements. Some new commitments cannot contractually be deferred, reduced or cancelled. The company consequently needs additional financing to meet forecast obligations in 2027, with uncertainty over whether capital will be available when needed on acceptable terms.
Nscale merits attention because large customer contracts and financing needs are growing together. Locking in equipment to fulfill orders reduces supply risk while bringing forward firm payment obligations. If construction and customer collections lag equipment payments, larger contracts can create greater near-term funding needs. For an enterprise considering long-term compute prepayments, orders and prominent customers demonstrate commercial traction. Delivered capacity and secured financing more directly determine protection of its own service commitments. Supplier assessment should examine order value alongside financing conditions still unmet before delivery.
Microsoft 365's outage shows that model availability covers only part of a workflow
The Microsoft 365 disruption beginning November 12 and continuing into the next day affected services including Outlook, Teams and Microsoft 365 Copilot. Microsoft said services were fully restored on November 14, according to the Associated Press. Workhorse announced that business disruption from the outage had forced it to postpone its third-quarter earnings release and conference call, originally scheduled for November 12. The disruption does not establish that all Azure model APIs were unavailable; service impact needs to be checked individually.
For enterprises connecting AI to email, meetings and internal documents, the dependency is clear. If a workflow needs material from an office system, a healthy replacement model cannot automatically restore access to that data. Models, data sources and business tools have different failure boundaries. Multimodel access addresses only some of them.
AgentsFlare's cross-channel routing and automatic failover support model-call availability. Business continuity also depends on access to data and tools, and whether a workflow can degrade safely. A control plane's value includes making those boundaries explicit: did the model fail, did the calling channel fail, or could an upstream business service no longer supply the material? Identifying the failure determines which alternative is relevant.
Models discount usage; compute providers negotiate commitments
This week's price divergence reflects different transactions. Fable reduces predecessor standard rates, Hydrogen pursues everyday execution, and Qwen uses limited-time Credit discounts to attract evaluation. GPU clouds sell capacity for specified locations, dates and durations. Nebius's higher new-deal prices and prepayment coverage show its bargaining position on those contracts. Both developments can occur together. An API price cut alone does not establish abundant compute supply.
Anthropic's proposed listing proceeds, CoreWeave's advance collections and Nscale's additional financing needs connect capital arrangements to service delivery. Enterprise AI autonomy requires the ability to compare real task costs and identify whether apparently separate suppliers share the same upstream dependencies. Keeping routing rules, budgets and auditing under enterprise control makes it possible to change suppliers when pricing, financing or delivery arrangements change—and keep the business running.