The agent bill moved from tokens to the whole system, just as power and financing became the binding constraints

Key takeaways

• Voice agents became execution systems: Google now keeps dialogue running while tools and deeper reasoning work in parallel.
• Infrastructure efficiency moved from component benchmarks to tokens per megawatt, with one measured Blackwell deployment lifting throughput 24% inside the same power budget.
• Capital markets supplied the counter-signal: CoreWeave sought $3B of convertible debt while debt tied to an Oracle-leased campus traded below par.
• The operating decision is to price agent work end to end, including the conversational layer, reasoning model, tools, recovery, and power.

By the numbers

Gemini 3.8 Live audio input

$0.005/min

Gemini 3.8 Live audio output

$0.018/min

Measured cluster token throughput

+24% — Lambda result reported by NVIDIA

CoreWeave convertible offering

$3.0B

Reported price of Project Jupiter loans

89-91¢

Flywheel arc · all-three

The token rate is no longer the cost of the agent; it is one line in a system bill.

• Google split real-time voice economics into independently metered input and output while background tools remain active.
• A measured Blackwell deployment produced 24% more cluster token throughput within the same power budget.
• CoreWeave and Oracle-linked financing show that infrastructure capacity is being repriced by capital markets, not only benchmarked by vendors.
• Cost per completed workflow is now the defensible procurement unit.

What this means

Architects should specify routing policy and the complete metering chain, not a favorite model. Automatic model selection now exposes an explicit cost-quality-latency policy, while live voice can keep a conversation open as tools and reasoning continue; see The Model Pulse for the model and pricing detail.

• Model routing became an explicit cost-quality-latency policy.
• Voice now remains active while tools and deeper reasoning execute in parallel.
• Domain post-training moved into a major system-of-record vendor's own trust boundary.

Sep 14

GitHub adds efficiency, balance, and intelligence tiers to automatic model routing

Sources GitHub Changelog

Sep 15

Google releases Gemini 3.8 Live and Live Extended Thinking through the Live API

Sources Google

Sep 15

Salesforce introduces Koa, a CRM reasoning model post-trained from NVIDIA Nemotron 3 Super

Sources Salesforce, NVIDIA

Sep 18

GitHub agents gain local Dev Containers and direct pull-request creation from agent sessions

Sources GitHub Changelog

What this means

Operators should make tokens per megawatt and performance under a site power cap acceptance metrics for new infrastructure. The measured Lambda result is stronger evidence than NVIDIA's forward Rubin multipliers, but both point to system-level power management becoming as important as accelerator peak throughput.

• Power-capped throughput is replacing peak silicon performance as the operating metric.
• Separate measured Blackwell results from vendor-projected Rubin multipliers.

Sep 15

NVIDIA reports Vera Rubin in production and publishes agentic throughput-per-megawatt claims

Sources NVIDIA

Sep 15

Lambda measures 24% more cluster token throughput under the same Blackwell power budget

Sources NVIDIA, Lambda

Sep 17

AMD argues for movable workloads across CPUs, accelerators, edge devices, and open interconnects

Sources AMD

What this means

Network buyers should test the fabric as part of a power-capped system rather than procure switching on headline bandwidth alone. The week's common design point is cross-layer control: compute, memory, fabric, cooling, and observability must expose enough telemetry to optimize one completed-work metric.

• Fabric evaluation now belongs inside the power-capped system test.
• Open interconnects matter when workload portability is an explicit resilience objective.

Sep 15

NVIDIA frames NVLink, Spectrum-X, ConnectX, BlueField, storage, and cooling as one power-managed factory

Sources NVIDIA

Sep 15

Cisco highlights cross-domain network, security, and observability context for agentic operations

Sources Cisco

Sep 17

AMD makes open interconnects and workload portability central to its AI infrastructure posture

Sources AMD

Frontier Labs detail

No frontier lab disclosed a financing or segment revenue figure during the window. Reuters reported investor scrutiny and possible release timing at Anthropic, but those are market signals rather than ledger inputs.

Capital in value

$0B

Revenue out value

Indirect

• Sep 19 · No qualifying disclosed transaction

Sources Reuters

Hyperscaler-Hosted detail

The week's category signal came from secondary trading rather than a new hyperscaler capital budget. Loans tied to an Oracle-leased New Mexico campus reportedly traded below par amid construction, environmental, and credit concerns.

Capital in value

$18B

Revenue out value

Indirect

• Sep 18 · Project Jupiter loans quoted below par · $18B outstanding

Sources Reuters citing Financial Times

Neoclouds detail

CoreWeave's financing is the category's fact home this week. The company launched $3 billion of convertible notes, an option for another $500 million, and an at-the-market program for up to 35 million shares, showing that capacity expansion remains capital intensive even with contracted demand.

Capital in value

$3B

Revenue out value

Indirect

• Sep 17 · CoreWeave convertible debt offering · $3.0B plus $500M option

Sources CoreWeave, Reuters

On-Prem / Hybrid detail

No government or enterprise on-premises program supplied a comparable financing number. AMD's portability argument and Salesforce's plan for controlled model deployment are product and architecture signals, not capital commitments.

Capital in value

$0B

Revenue out value

Indirect

• Sep 17 · No qualifying disclosed transaction

Sources AMD

Signal score 4/5

Lambda raised cluster token throughput 24% inside the same power budget by running 19 nodes where 16 full-power nodes were normally allocated.

This is the strongest operating evidence in the window because it names the before and after configuration. Buyers should request the same power-capped test on their workload.

Sources

NVIDIA report of Lambda measurement

Signal score 3/5

Google's Gemini 3.8 Live can continue a conversation while tools and deeper reasoning run in the background.

The capability is shipped through the Live API, but outcome quality and interruption behavior still need workload-specific testing.

Sources

Google launch posts

Signal score 2/5

Vera Rubin delivers up to 30 times more agentic throughput per megawatt than GB300 NVL72.

The number is vendor-reported and workload-specific. Treat it as a test target, not a capacity-plan input, until independently reproduced.

Sources

NVIDIA vendor benchmark using AgentX

Signal score 2/5

Koa produces three times fewer errors than leading models on CRM actions.

The benchmark, dataset, and comparison set are vendor-controlled. Pilot availability is real; the multiple is not yet procurement-grade evidence.

Sources

Salesforce CRM benchmark

Abductive · 76% confidence

AI procurement is moving from model selection to system-yield contracting.

Steel-man: These are vendor-defined meters and one customer measurement, not a standardized contracting unit. The claim is bounded to procurement direction, not current market practice.

• Google split live-agent audio into input and output meters while background reasoning and tools continue.
• GitHub exposed cost, quality, and latency as selectable routing policy.
• NVIDIA and Lambda measured completed token throughput under a fixed power budget.

Sources Google Gemini 3.8 Live developer pricing · GitHub auto model tiers · NVIDIA and Lambda power-capped measurement

Inductive · 69% confidence

Infrastructure credit risk is becoming an operational architecture input rather than a finance-only concern.

Steel-man: Both financings may be issuer- or project-specific and demand remains strong. The inference is about diligence scope, not a sector-wide credit event.

• CoreWeave sought $3 billion of convertible debt plus an equity-sale facility to fund operations.
• Loans tied to an Oracle-leased campus were reported below par amid project-specific concerns.
• Hardware vendors simultaneously shifted their headline metric to useful work inside a fixed megawatt.

Sources CoreWeave financing · Oracle-leased campus debt

Hypothesis 1 · Supported

Software demand pulls hardware and network investment forward.

Parallel voice reasoning increases the number of simultaneous model, tool, and memory paths, while infrastructure vendors answer with system-level throughput optimization.

Counter-evidence: No buyer disclosed incremental capacity ordered specifically for Gemini Live.

Sources Gemini Live launch · NVIDIA AI factory result

Hypothesis 2 · Supported

Hardware efficiency expands economically viable AI demand.

Lambda's measured 24% throughput gain inside the same power budget directly lowers the infrastructure consumed per token.

Counter-evidence: One Blackwell deployment does not establish portability across models or facilities.

Sources Lambda DSX measurement

Hypothesis 3 · Supported

Networking becomes a first-order limiter as AI systems scale.

The week's vendor architectures treat fabric, memory, storage, and cooling as one controlled system rather than separable components.

Counter-evidence: No in-window outage or benchmark isolated networking as the binding bottleneck.

Sources NVIDIA factory architecture · AMD open infrastructure posture

Hypothesis 4 · Strained

Capital follows visible utilization and contracted demand.

CoreWeave accessed large financing, but below-par project debt shows that contracted demand no longer neutralizes construction, environmental, and concentration risk.

Sources CoreWeave notes · Project Jupiter loans

Hypothesis 5 · Untested

Open ecosystems gain when switching costs become material.

AMD emphasized open interconnects, but the window supplied no comparable buyer migration or audited share shift.

Sources AMD infrastructure essay

Inductive · 2 weeks observed

Agent cost decomposes into independently managed system meters

Next expectation: A major provider will publish a worked multi-meter agent cost example before year end.

• W37: GPT-Live-1 split voice from separately billed reasoning.
• W38: Gemini Live split audio input and output while tools run in parallel.

Inductive · 2 weeks observed

Power-capped throughput replaces peak component performance

Next expectation: At least one infrastructure launch next week will headline useful work per watt or megawatt rather than peak FLOPS.

• W37: long-context releases led on cache bytes per token.
• W38: Lambda reported tokens per second and performance per watt inside a fixed power budget.

Next two quarters

Voice agents continue talking while background tools and reasoning execute.

FinOps must attribute one user interaction across several concurrent meters and failure domains.

Who moves

CIOs, application architects, and AI platform teams

Next 12 months

AI campus debt trades below par while new capacity still seeks billions in financing.

Workload portability and staged capacity commitments become credit-risk mitigations, not only architecture preferences.

Who moves

Cloud buyers, lenders, neoclouds, and infrastructure vendors

Frontier lab cash runway at current burn

The required inputs remain unaudited and absent. GPT-Live-1 list pricing and the Agents API's managed-service terms are product and commercial signals, not financing inputs, and no lab published a raise or a revenue figure this week.

Hyperscaler AI capex to disclosed AI revenue ratio

Oracle's filing supplies a real capital-expenditure number and a real backlog number but no AI-segment revenue, which is the denominator the method requires. No hyperscaler disclosed AI-segment operating margin this week, so the ratio stays unpublished rather than estimated.

CoreWeave contracted revenue backlog

Backlog remains a filed stock value awaiting the next quarter. The week's neocloud financings, an undrawn term loan and a convertible note with an investor put, are category structure signals and cannot be substituted into this series.

NVIDIA quarter-over-quarter data center revenue

The in-window NVIDIA news is an Australian capacity aggregation across eight operators and a third-party accelerator joining NVLink Fusion. Neither is a revenue disclosure, and TSMC's record month is a supplier datapoint that cannot be substituted into this series.

Open-weight to closed-model capability gap on coding

This is a measurement failure rather than a capability result. Two open checkpoints shipped in the window, both explicitly cheaper and smaller rather than stronger, and one carries vendor-reported benchmarks with no independent evaluation. See Model Pulse for the lineage read.

Sovereign AI program commitments

The method counts government-funded national compute programmes only. This week's two candidates fail it for different reasons: an at-least-EUR-13B corporate investment plan with a power purchase agreement is private capital, and an up-to-two-gigawatt national target is an aggregation of eight commercial operators' pipelines.

PJM capacity auction clearing price

The prior threshold, a second consecutive auction clearing at the cap, is already satisfied three times over and can no longer move: 2026/27 cleared at $329.17, 2027/28 at $333.44 and 2028/29 at $325.00, each at the approved cap. PJM's own no-cap-or-floor simulation for 2028/29 reports $554.72 per MW-day for the RTO and $776.69 for ComEd, versus the $325 capped result, with $29.7 billion of simulated cleared value against $16.4 billion actual. The simulation does not isolate the cap as the sole cause, but it quantifies how much higher PJM's model clears without it. The in-window development is regulatory rather than price-setting: proposed large computational load ride-through, ramp-rate, telemetry and remote-disconnect requirements were presented on September 8, targeting a federal filing in November 2026. If mandatory curtailable status attaches to gigawatt-scale loads, it changes what capacity payments are actually buying.

Time from interconnection request to energization

No comparable queue-duration update was published. A Texas docket fight over forfeiture terms on a 75-megawatt interconnection rule is evidence of queue scarcity and of the cost of holding a position in one, but it is not a duration measurement.

Cost per task, frontier reasoning model

The cost spread is the most durable finding to survive the index revision and is a genuine procurement fact: 57% lower cost for the same rounded composite score. It is reported here as two point observations rather than as movement, because the composite they price changed underneath them.

Custom silicon share of hyperscaler AI compute

The week strengthens the directional case for custom inference silicon on two counts, a multi-generation agreement and a third-party accelerator entering the dominant scale-up fabric, while supplying no unit volumes, deployed utilisation, or mix disclosure from which a share could be computed.

Software · 43% confidence

At least one major voice-agent provider publishes an end-to-end worked cost example that includes voice, reasoning, and tool execution by December 31, 2026.

ID

p112-live-api-cost-disclosure-dec31

Deadline

By December 31, 2026

Trigger

Hit only if a first-party pricing or documentation page shows all three cost components in one worked workflow.

Hardware · 31% confidence

A major server, accelerator, or cloud vendor publishes a customer procurement template using tokens per megawatt as an acceptance metric by December 31, 2026.

ID

p113-power-capped-rfp-dec31

Deadline

By December 31, 2026

Trigger

Hit only with a public first-party RFP, reference architecture, or acceptance guide naming tokens per megawatt.

Capital · 84% confidence

CoreWeave completes at least $3 billion of the convertible note offering announced September 17 by November 30, 2026.

ID

p114-coreweave-financing-close-nov30

Deadline

By November 30, 2026

Trigger

Hit only if a CoreWeave filing states gross proceeds of at least $3 billion from the announced notes.

Software · 67% confidence

Salesforce makes Koa generally available in at least one US region by March 31, 2027.

ID

p115-koa-ga-mar31

Deadline

By March 31, 2027

Trigger

Hit only if Salesforce documentation marks Koa generally available, not pilot or beta, in a named US region.

Networking · 37% confidence

A major AI networking vendor adds workload-level token-throughput correlation to a generally available fabric telemetry product by March 31, 2027.

ID

p116-fabric-power-telemetry-mar31

Deadline

By March 31, 2027

Trigger

Hit only if public product documentation joins network telemetry to model token throughput for a named workload.

Pending · Capital

A publicly traded operator other than Oracle discloses, for a specific reporting period, both a megawatt capacity figure delivered or placed in service and a unit count of AI accelerators delivered, by March 31, 2027.

ID

p105-second-operator-delivered-capacity-mar31

Confidence

62%

Deadline

By March 31, 2027

Trigger

Hit only if an earnings release, 10-Q, 10-K, or call transcript from an operator other than Oracle states both a megawatt capacity figure delivered or placed in service and a count of accelerators delivered or deployed, for a named period. A backlog, contracted-capacity, or announced-pipeline figure is a miss.

Pending · Power

Fortum announces an approved investment decision covering at least EUR 300 million of the EUR 700 million of Loviisa life-extension capital expenditure currently disclosed as pending, by June 30, 2027.

ID

p106-loviisa-fid-jun30

Confidence

44%

Deadline

By June 30, 2027

Trigger

Hit only if a Fortum release, interim report, or annual report states an approved investment decision of at least EUR 300 million for the Loviisa life-extension programme. Reaffirmation of the programme, or approval below EUR 300 million, is a miss.

Pending · Software

A major model provider or evaluation publisher ships a machine-readable runtime or harness manifest that ties a published score to a reproducible configuration, by March 31, 2027.

ID

p107-runtime-manifest-mar31

Confidence

38%

Deadline

By March 31, 2027

Trigger

Hit only if public documentation or a results artifact contains a versioned configuration identifier naming at least harness or adapter version, tool permissions, context persistence policy, and retry budget, which is the same four-element bar this issue's pattern watch sets for the same event. Prose disclosure of methodology changes, or publishing two harness results side by side without a configuration identifier, is a miss.

Pending · Power

PJM's Section 205 filing on large computational loads is docketed by December 31, 2026 and carries a telemetry or remote-disconnect requirement, not merely a ride-through envelope or a ramp-rate limit.

ID

p108-pjm-large-load-filing-dec31

Confidence

58%

Deadline

By December 31, 2026

Trigger

Hit only if a FERC filing by PJM docketed on or before December 31, 2026 includes a telemetry or remote-disconnect requirement applicable to large computational loads. A filing carrying only a voltage or frequency ride-through envelope or only a ramp-rate limitation is a miss, as is a further stakeholder presentation or any slip past year end.

Pending · Networking

At least two distinct vendors announce 1600ZR-conformant coherent pluggable optics products by June 30, 2027.

ID

p109-1600zr-two-vendors-jun30

Confidence

66%

Deadline

By June 30, 2027

Trigger

Hit only if product announcements or datasheets from two distinct vendors cite conformance or compliance with the OIF 1600ZR Implementation Agreement. Demonstrations, interoperability plugfests, and roadmap statements without a named product are a miss.

Pending · Software

OpenAI's managed Agents API supports zero data retention or a non-US data residency option by March 31, 2027.

ID

p110-agents-api-residency-mar31

Confidence

29%

Deadline

By March 31, 2027

Trigger

Hit only if OpenAI documentation states zero-data-retention eligibility or a non-US data residency option for the managed Agents API. A self-hosted sandbox, a roadmap statement, or residency support on other OpenAI surfaces but not the Agents API is a miss.

Pending · Hardware

Qualcomm or its counterparty discloses a named service date, first-deployment date, or unit volume for the multi-generation custom AI inference agreement, by March 31, 2027.

ID

p111-custom-inference-service-date-mar31

Confidence

33%

Deadline

By March 31, 2027

Trigger

Hit only if a filing, release, or call transcript from either party states a calendar service or first-deployment date, or a unit or megawatt volume, for the agreement. Restating the transaction value, the warrant structure, or a generational roadmap without a date or volume is a miss.

Cumulative record

Predictions made

116

Resolved

57

Outcomes

23 hit · 15 partial · 19 miss

Hit rate (partial = half)

54%

Brier score (0 = perfect)

0.200

Overdue, unresolved

0

Hit · 72% called

NVIDIA files exhibits with the 10-Q for the quarter ended July 26, 2026 that translate the SB Energy PORTS-Pike residual-value guaranty into a per-quarter contingent-obligation disclosure and identify the OpenAI affiliate as tenant, by October 31, 2026.

NVIDIA filed the Form 10-Q for the quarter ended July 26, 2026 on August 26, 2026 — inside the window. It satisfies all three trigger elements: guarantees 'capped at a total of $105 billion' with an exposure table of $3.5B AI-cloud guarantees plus $105.0B SB Energy for $108.5B total; effectiveness conditioned on SB Energy satisfying applicable ready-for-service conditions as each of nine phases is placed in service from fiscal 2029; and the tenant identified as 'an affiliate of OpenAI Group PBC' at the PORTS Technology Campus in Pike County, Ohio. Exhibit 10.1 is the Form of Residual Value Guaranty.

Deadline

By October 31, 2026

Partial · 80% called

Aggregate 2026 hyperscaler capex revises upward by 10% or more from the $700B baseline.

Q1 prints (MSFT $190B, GOOG $180-190B, META $125-145B, AMZN $200B reaffirmed) take 2026 aggregate to $695-725B (+77% YoY) vs the $700B W17 baseline. At/near baseline; +10% revision (~$770B) plausible by Q2 print. Score moves to hit if Q2 takes aggregate above $770B.

Deadline

By October 31, 2026

Hit · 43% called

Z.ai publishes GLM-5.3 weights to Hugging Face by September 15, 2026, closing the two-week window promised at the model's August 14 announcement.

Z.ai published the full 753B-parameter GLM-5.3 weights to Hugging Face at zai-org/GLM-5.3 on August 27–28, 2026 — in-window and inside the trigger's September 15 window, distinct from GLM-5.2 — after GLM-5.3-Flash MIT weights landed Aug 26. The material nuance is licensing, not availability: GLM-5.3 ships under a bespoke GLM-5.3 license rather than MIT, requiring Z.AI security review before commercial use by any Model-as-a-Service operator whose group revenue exceeds $10B over any 12 consecutive months.

Deadline

By September 15, 2026

Hit · 66% called

An independent benchmark finds Gemini 3.6 Flash at least 12% cheaper per completed agentic task than Gemini 3.5 Flash by August 31, 2026.

Artificial Analysis measured Gemini 3.6 Flash at $0.50 average cost per completed agentic task versus $0.59 for 3.5 Flash — a 15% reduction, above the 12% cheaper-per-task bar — before Aug 31.

Deadline

By August 31, 2026

Sources Resolution evidence

Hit · 84% called

DeepSeek V4's official GA pricing does not reset the ultra-cheap floor: off-peak deepseek-v4-pro output pricing stays at or above ¥6 (~$0.85) per MTok through August 31, 2026 — the kill-condition test for this issue's price-band-convergence claim.

DeepSeek's official API pricing page kept GA deepseek-v4-pro off-peak output at $1.98/MTok (~¥14+) through Aug 31 — well above the ¥6 (~$0.85)/MTok ultra-cheap floor the trigger set as the kill condition.

Deadline

By August 31, 2026

Sources Resolution evidence

Hit · 64% called

At least one major agent platform (OpenAI, Anthropic, GitHub, or Cursor) ships product-level per-task or per-harness cost telemetry or routing controls — beyond session budget caps — by August 31, 2026.

Cursor shipped Cursor Router in July 2026 with Auto Balance/Intelligence routing controls and published measured cost-per-commit figures ($4.63–$6.76) from live traffic — product-level harness routing and cost telemetry beyond session budget caps.

Deadline

By August 31, 2026

Sources Resolution evidence

Sep 21-25

Independent Gemini 3.8 Live latency and interruption tests

The shipped rate card is useful only when paired with measured turn latency, tool-call delay, and interruption recovery.

Sep 21-30

CoreWeave convertible pricing and closing

Coupon, conversion premium, hedging cost, and final proceeds will show what capital markets charge for the next unit of neocloud capacity.

Sep 21-Oct 2

Project Jupiter credit and permitting response

Any lender, county, or Oracle disclosure could distinguish temporary syndication pressure from a material construction risk.

Oct 2026

Koa pilot evidence and open-beta timing

Named pilots need workload definitions and error baselines before the vendor's three-times-fewer-errors claim can inform procurement.

• Authored from verified public sources dated September 14-19, 2026, with primary vendor material used for product claims and Reuters used for financing and market reports.
• No matured W37 prediction deadline fell inside the observation window; all seven prior predictions remain pending.
• The fact home for CoreWeave financing is Capital Flow; other publications use only short cross-references.