Anthropic's newly disclosed July-end ARR of $65B came in below market expectations, as YipitData's more optimistic estimate had already reached $67B by the end of June. Some market participants speculate that the shift from a weekly to a monthly run-rate reflects expectation management ahead of an IPO.
02 · GPU PRICINGA100/H100/H200/B200 rental prices changed +0.0%/+0.7%/+4.9%/+0.5% over the past 7 days. H100 is up 29.8% YoY and near its one-year high: immediate availability still matters more than theoretical generation efficiency.
03 · SUPPLYEpoch satellite tracking shows 11.55GW of operational IT capacity and 23.24GW of future pipeline at 2Q26; operational capacity is modeled to reach 13.05GW in 3Q26, up 1.50GW q/q.
04 · OPENROUTERPaid open-weight models generated 41.4T tokens this week, up 7.4% w/w. Within modeled open-weight revenue, Z.ai holds 34.6%, Kimi 30.0% and DeepSeek 16.6%.
05 · CREDITCoreWeave DDTL 5.5 is a $2.6B facility priced at SOFR +550bp and supported by unrated take-or-pay customers.
07 · MODELSGLM-5.3 approaches—but does not exceed—Kimi K3 in independent testing; its smaller model and lower task cost improve the economics, while cybersecurity tasks stand out as a notable strength.
NVIDIA server prices may rise more than 15%; the direct impact is higher system procurement cost, while contract pass-through determines the project-return effect.
Nebius secured more UK capacity; delivery improves, but revenue still depends on fit-out, GPU installation and customer activation.
OpenAI expanded Zero Data Retention support, removing a key barrier for regulated enterprise inference and helping it gain enterprise share versus Fable 5, which still does not support ZDR, while increasing model-distillation risk.
OpenAI slowed frontier scaling on cyber risk; launches may slip, while security monitoring itself consumes more compute.
Z.ai believes the largest marginal opportunity for model improvement now lies in post-training; GLM-5.3 improved with model size unchanged.
01 / REVENUE MOMENTUM
Track the latest disclosed annualized revenue and estimate where each company stands today.
Revenue run-rate
Current projected ARR
+$8.5B estimated since July month-end
Latest reported ARR
$65.0BJul 31, 2026Last-month MoM
+17.6%June → JulyReported as >$65bn; tracker uses $65bn lower bound
Revenue run-rate
Current projected ARR
+$2.9B estimated since July month-end
Latest reported ARR
$40.0BJul 31, 2026Last-month MoM
+9.9%June → JulyCompany internal communication reported by Axios
Revenue momentum is accelerating while both legacy and frontier GPU rental indices remain firm.
Anthropic's latest monthly growth is running ahead of OpenAI's, pointing to broadening model demand rather than a single-company cycle.
A100 and H100 are both positive year-on-year. Near-term availability is outweighing theoretical efficiency gains.
H100 trades near its one-year high, suggesting the market continues to pay for immediately deployable capacity.
“Latest reported ARR” is the public anchor. “Current projected ARR” is a simple timing nowcast, not company guidance. It assumes the previous month's growth accrued evenly across the current month.
02 / GPU RENTAL PRICING
Track current neo-cloud on-demand rental pricing and where each GPU sits within its one-year range.
H100 is up 29.8% year-on-year and trades close to its one-year high; H200 has the strongest 30-day move. Older Hopper capacity is not clearing like a rapidly depreciating commodity.
Nebius reported Q2 deals above $20m ACV/MW, with four landmark deals at $20–25m/MW. Early-Q3 short-term capacity reached a $40–50m/MW pricing opportunity. These are new-deal economics, not current company ARR divided by total power.
Source: Silicon Data. Limited summary of standard GPU indices viewed on Aug 23, 2026; index observations are as of Aug 22, 2026. B300 is omitted because no standalone Silicon Data index was available.
03 / COMPUTE SUPPLY
Track operational IT capacity, future development pipeline and the pace at which planned supply is becoming usable compute across major U.S. data centers.
Epoch-tracked operating capacity
Tracked construction and future development pipeline
Expected operating capacity
Expected q/q capacity addition
Source: Epoch AI satellite imagery tracking. SpaceXAI is anchored to an SEC-reported 1.0GW nameplate compute draw at 1Q26, then extended with Epoch's modeled Colossus II staging. Meta's 1.94GW is more model-dependent: company disclosures confirm Prometheus' 1GW target, but not every site-level quarter-end estimate. “Future pipeline” includes tracked future capacity and should not be read as a clean under-construction-only measure. 3Q26–4Q27 are model periods.
05 / COMPUTE CREDIT
Track how customer credit support, collateral structure and refinancing conditions translate into neo-cloud funding costs.
GPU / contract-backed term financing
CCIR tracked amount
Floating-rate funding cost
Take-or-pay customers
Spread over SOFR. These facilities are not identical; the comparison is a directional read on how contract quality changes marginal funding cost.
DDTL 5.5 provides a current marginal funding-cost anchor for projects backed by unrated take-or-pay customer contracts: SOFR plus 550bp before fees and other structural costs.
Source: CCIR Compute Credit Tracker and Credit Reference Series; selected quoted values as of 18 Aug 2026. Full time series and ledger are not reproduced.
04 / MODEL USAGE
Weekly token traffic shows where inference demand is actually routing across models and providers.
Selected paid models; free variants excluded
Theoretical, at modeled blended price
Directional only; direct API traffic excluded
Exact trailing 7-day volume versus prior 7 days
90% cache hit rate · 10% uncached input · input/output tokens = 100:1. Blended $/MTok = input share × (90% × cached-input price + 10% × input price) + output share × output price. Theoretical weekly revenue = weekly tokens × blended $/MTok.
Source: OpenRouter (openrouter.ai/rankings), Data API as of Aug 23, 2026; usage windows Aug 9–15 and Aug 16–22. Licensed under CC BY 4.0. Free variants, NVIDIA Nemotron and unidentified free models are excluded. Open-weight and closed-model value shares are normalized only within their own samples and must not be compared with each other; closed-model direct API traffic is not captured.
06 / WEEKLY INTELLIGENCE
Memory inflation is pushing up systems based on Vera Rubin and Grace Blackwell, with increases expected on early-2027 shipments.
The direct read-through is higher system procurement cost. The effect on neo-cloud project returns depends on contract pricing and the operator's ability to pass those costs through.
Nebius will lease NVIDIA-powered high-density capacity at Vantage's Newport campus in the UK's South Wales AI Growth Zone.
Positive for Nebius delivery capacity, but the revenue timing still depends on fit-out, GPU deployment and customer activation.
Eligible API customers can use frontier models without OpenAI retaining prompts or responses after processing.
Reduces a major blocker for regulated enterprise workloads and can expand high-value inference demand. In comparison, Anthropic supports ZDR for Opus 5 but not Fable 5; mandatory retention may help detect or deter model-distillation attempts against Fable 5.
OpenAI cited critical cyber-capability risks and higher security requirements; monitoring overhead is estimated at roughly 20% of affected inference compute.
Near-term launch timing risk rises, while security monitoring itself adds compute demand and raises the cost of frontier development.
Z.ai's current view is that parameter count is not the only training variable worth scaling. GLM-5.3 kept the same base architecture, total parameters and activated parameters as GLM-5.2, then improved capability by spending one month scaling long-horizon environments and reinforcement learning. The company sees the largest current marginal opportunity in post-training, while pre-training, mid-training, model size and compute per forward pass remain future scaling directions.
Z.ai believes the largest marginal opportunity for model improvement now lies in post-training. A smaller model can therefore approach the capability of a much larger peer without matching its training or serving cost; compute spending may shift toward long-horizon environments and reinforcement learning rather than parameter scaling alone.
Sources: official company disclosures and filings first; specialist datasets and major financial media are used for confirmation.
07 / NEW MODEL WATCH
A rolling two-week comparison of official evaluations, outside reaction and the investment implication behind each launch.
Artificial Analysis v4.1.1 aggregates nine reasoning, knowledge, coding and agentic evaluations. GLM-5.3 now scores 60, level with Kimi K3 and one point behind Grok 4.6.
Higher is better; the x-axis runs from expensive on the left to cheaper on the right. Each connected point represents a different reasoning-effort setting for the same model.
Z.ai
Z.ai reports a 50% improvement over GLM-5.2 on its internal coding benchmark and open-model SOTA on several public agent benchmarks.
Independent Chinese testing finds GLM-5.3 close to Kimi K3 in coding, but not above it overall. GLM-5.3 uses materially fewer steps and costs only one-third to one-half as much per task as K3 or Qwen3.8; the trade-off is weaker edge knowledge and less self-testing.
能力接近但实测没有超过Kimi K3；考虑到模型参数更小、单任务成本更低，GLM-5.3的经济性更好。
总体接近Kimi K3；开发步骤更少、并发工具调用更积极，同任务成本约为K3或Qwen3.8的1/3–1/2。
审美、边缘知识和上下文召回弱于K3；自测偏少，偶发忽略原始要求造成的功能遗漏比交互细节问题更严重。
通用推理输出量较前代上涨约24%，已进入偏低效区间；但在Agent任务中规划较克制，max档全程使用仍具实用性。
GLM-5.3 sits close to Kimi K3 on intelligence while costing less per task. It does not exceed K3 on the independent score, but it occupies the more attractive cost position.
Capability is close to Kimi K3 but does not exceed it in independent testing. Because GLM-5.3 reaches that level with a smaller model and much lower task cost, its economic position is stronger than the raw score alone suggests.
Google positions Flash as its high-throughput, low-latency model tier. Gemini 3.7 pairs stronger reasoning and agent performance with service speed approaching 200 output tokens per second.
Independent Chinese testing finds a large jump from Gemini 3.5 Flash: coding reaches the high-usable tier, instruction following improves without sacrificing initiative, and the Low setting nearly doubles reasoning performance with only a modest token increase.
Gemini重新回到北美头部模型序列，3.7 Flash是这一轮最明确的翻身之作。
较3.5 Flash大幅跃升至高可用水平；既保留主动补充交互、UI和常见功能的能力，也显著改善了指令遵循。主要短板是大型存量工程探索不足，偶尔忽视向前兼容。
Low档平均思维Token约从3.5 Minimal的6K增至10K，但推理性能接近翻倍；High档平均约26K，明显低于同级Qwen3.8-Max和DeepSeek V4 Pro。
接近200 tok/s的输出速度在相同性能区间缺少对手。速度、较低任务成本与高可用Agent能力的组合，比单项榜单排名更具商业意义。
Gemini 3.7 Flash is the family's comeback release. It does not lead the frontier on raw intelligence, but its combination of speed, low reasoning cost and now-credible agent performance makes it one of the strongest commercial inference products.
SpaceXAI
The official evaluation puts Grok level with GPT-5.6 Sol on the Artificial Analysis composite and close to Fable 5 on agentic knowledge work.
Independent coverage views the release as a return to the frontier tier, while noting that OpenAI and Anthropic retain a narrow lead on several demanding tasks.
A credible frontier alternative, not a decisive leader. The $2 input / $6 output pricing is strategically more important than small benchmark gaps.
Meta
Meta positions the 30B multimodal model as an always-on local agent with reliable tool use, long-horizon workflows and an Apache 2.0 license.
Developer tests broadly validate the deployment claim: users report that the official quant fits on 24GB-class hardware and performs especially well in tool calling.
Not a frontier-cloud substitute. Its significance is shifting some agent inference to local hardware and lowering the cost floor for private, always-on workflows.
Method: developer benchmarks are not directly comparable unless the harness, reasoning budget and tool scaffold match. Official scores are shown as launch evidence—not treated as independently verified facts. X and community feedback are used to identify real-world strengths or failures, never as the sole performance conclusion.