Stripe buys OpenRouter for over $7B, OpenAI keeps zero retention and halts training for two weeks, Anthropic's run rate reaches $65B

Over the past week the dividing line between model vendors and enterprises moved onto data. On August 19, OpenAI committed to keeping Zero Data Retention for enterprise customers on its frontier models and previewed a mechanism that detects abuse without touching customer content. Over the same stretch, Anthropic has required 30-day retention on Claude Fable 5 and Mythos 5 with no opt-out, and said outright in a risk report that this will be unpopular with customers and poses real risks to the business. The two diagnose the problem identically and choose opposite commercial answers, and Ramp's payments data shows enterprises have already started voting along that line. Three more moves pushed in the same direction the same week: Stripe paid more than $7 billion for the model aggregation gateway OpenRouter, roughly 5.4x its valuation three months earlier; OpenAI paused reinforcement learning training for two weeks on safety grounds and, for the first time, put a public number on what safety monitoring costs in compute — about a fifth of the inference compute being monitored; and Anthropic's annualized revenue passed $65 billion at the end of July, with a pre-IPO revolving credit facility set to exceed $10 billion. In hardware, Groq took NVIDIA's money and turned to building an NVIDIA cloud while Etched doubled its valuation to $21 billion in a month — dedicated inference silicon took two opposite paths in the same week.

What exactly $7 billion bought

On August 16, Bloomberg reported that Stripe had finalized an agreement to acquire OpenRouter for more than $7 billion; CNBC confirmed on August 19. Some reports put the figure at about $7.5 billion with $1.5 billion allocated to the founders — sourced from people briefed on the deal, not disclosed by either party. OpenRouter is a single interface fronting more than 400 models, with centralized billing and request routing, serving roughly 8 million users. It closed a Series B at $1.3 billion just three months ago in May, making the purchase price about 5.4x that mark. Stripe CEO Patrick Collison gave the rationale as helping customers maximize profitability by routing requests intelligently and spending tokens efficiently.

A payments company buying the model aggregation layer says that layer is now worth an acquisition price, and the pricing logic resembles payments: the transaction flow itself is worth money, and whoever sits on the path that flow must take can turn metering, settlement, and risk control into products. For the past two years, unifying interfaces across model vendors was dismissed as a thin layer of glue code that any vendor's official SDK could flatten. A 5.4x mark-up in three months says the market has changed its mind, and the reasoning is not hard to follow: model price lists change faster than enterprise procurement cycles. Gemini 3.7 Flash's $0.75/$3.75 promotional rate expires December 31; DeepSeek moved to peak and off-peak Beijing-time billing on August 17; GPT-5.6 Luna was cut 80% three weeks after launch. At that cadence, which model you call is a decision you can revise often, while which layer your traffic passes through is hard to change.

When the aggregation layer changes owners, what the layer is for changes with it. OpenRouter grew up serving developers and AI-native applications, selling a unified interface, centralized billing, and routing; folded into a payments infrastructure company, its capabilities will most likely grow along settlement, metering, and merchant risk. That direction only partly overlaps with the questions enterprises have to answer internally. Which team, acting for which user, called which model and which tool, spent how much, whether that request crossed an internal access rule, and whether the whole chain can be pulled up afterward — those questions have to line up with the company's own identity system and compliance definitions, and an external billing gateway cannot answer them. Hold access rules, per-endpoint and per-team cost attribution, and request-level audit in your own hands, and however price lists move or the aggregation layer changes hands, what changes is configuration.

Zero retention or thirty days: who holds the data

On August 19, OpenAI published "Offering Zero Data Retention for frontier models," confirming that frontier models remain available to eligible API customers under Zero Data Retention, and previewing a mechanism called Private Safety Processing. The ZDR promise is that prompts and model responses are not retained once a request is processed, that OpenAI personnel cannot review them, and that enterprise customer data is not used for training absent explicit opt-in. The new mechanism targets a different problem: some risks only show up when several interactions are read together — the same party asking about weaknesses in a piece of software in one conversation, then asking about remote access and detection tooling in another. The approach keeps customer content on customer-controlled infrastructure, or on OpenAI infrastructure encrypted with customer-held keys; automated systems identify patterns across sessions on top of it and return only an alert category and severity, with the underlying text never reaching OpenAI personnel. It is in test with early customers, with broader rollout and a technical white paper in September. The scope is drawn clearly: enterprise and API customers, with consumer subscriptions excluded.

At the other end is Anthropic. Since the June 9 launch of Claude Fable 5 and Mythos 5, both have been designated covered models, with prompts and outputs retained for 30 days across first-party and third-party platforms, no opt-out, regardless of what prior contracts said. Anthropic's language in its August risk report is unusually direct: the decision will be unpopular with customers who have come to expect zero retention and poses real risks to the business, especially if competitors do not follow, but that window is essential to detect and prevent sophisticated attacks spanning multiple requests. Both diagnose the problem the same way — cross-session risk is real. The disagreement is over who should keep that context.

The split already shows in procurement data. Ramp's payments data across more than 70,000 US businesses shows Anthropic overtaking OpenAI for the first time in May at 41% to 39%, reaching roughly 44% to 40% in July, while OpenAI has grown faster among these customers quarter-to-date. A Ramp economist attributed Fable 5's weaker-than-expected adoption to two things: price and data retention requirements. The data can only say so much: the sample skews toward technology companies, does not cover large enterprises using other spend-management tools, and Ramp gave percentages rather than dollar figures. What is certain is that the market keeps expanding — the share of these customers paying for AI passed half in March and reached nearly 56% by July.

Data retention used to be a line in a contract annex; now it is a hard constraint at selection time. In regulated sectors such as financial services, healthcare, and law, whether a vendor may retain content is decided by external obligation rather than procurement preference; a flagship model that cannot run under zero retention will not get into those use cases however capable it is. The trouble runs inside a single company too: the same rule has to be implemented once against OpenAI's zero-retention deployment, once against Anthropic's 30-day covered models, and once against self-hosted open-weight models — and not one of the three aligns on retention window, key custody, or how an alert is defined. Making one set of redaction and access rules effective across every model and settling request-level audit in one place is the only way this holds together in a multi-model environment. For enterprises running AI traffic on AgentsFlare, a vendor retention policy change means adjusting routing and access policy once, with no code changes across business lines.

A fifth of the compute goes to watching the model

On August 18, OpenAI published "Pacing model development in an era of cyber-critical capabilities," explaining how two developments over the preceding weeks changed its own research cadence: the July OpenAI-Hugging Face incident, and preliminary evidence on August 7 that its upcoming Astra model may meet the Critical cybersecurity capability threshold under its Preparedness Framework. The company temporarily slowed scaling and paused reinforcement learning training for two weeks on the latest models intended for deployment, using the time to harden research environments, run red-team exercises, and expand monitoring coverage. Its largest planned frontier RL run remains on hold, replaced for now by smaller-scale training and evaluations to validate safeguards and establish alignment evidence. After the Hugging Face incident, frontier model inference in research clusters that could execute code or reach the internet through tools was halted outright, then restored workload by workload.

The new monitoring stack is multistage: activation classifiers running at every sampled token handle first-pass screening and escalate concerns to higher-compute automated investigators, which examine tool actions, available reasoning, and the full activity sequence for unauthorized access, data theft, destructive behavior, and attempts to defeat safeguards. The target is an alert within 30 minutes of concerning activity surfacing; if a critical security boundary is implicated, safety, security, and research teams are paged at once, and if they cannot conclusively rule out a false positive within 30 minutes, the activity is expected to pause. The scope covers all tool-enabled RL training and evaluation for models at Sol capability or above; after the August 7 determination, all tool-enabled Astra inference was included as well.

The number worth remembering is that percentage. OpenAI estimates monitoring overhead at roughly 20% of the inference compute being monitored, with substantial variation across training and evaluation workloads. This is the first time a frontier vendor has publicly quantified how much compute safety monitoring consumes. That overhead hits price first. Model-layer price cuts over the past year came mainly from inference efficiency and hardware iteration, and if high-capability tiers must run permanently dragging a fifth of their compute in monitoring, that cost eventually walks into the price list, likely widening the gap between high and low tiers rather than shifting the whole curve down as previous rounds did. It hits delivery too. The company wrote plainly that the new security standards imposed great cost on frontier research and slowed it down, and that a significant number of workloads remain paused until migration completes. Putting an unreleased model tier into next year's product roadmap now means reserving a stretch of time the vendor itself cannot pin down.

A $65 billion run rate, and a $10 billion credit line

On August 17, Bloomberg reported that Anthropic's annualized revenue passed $65 billion at the end of July, up from $47 billion in May and $9 billion at the end of last year — roughly $18 billion added in two months. Preliminary second-quarter revenue disclosed at the same time exceeded $11.5 billion, against $787 million a year earlier. The figures come from a routine investor update, not audited public disclosure. The Financial Times reported that investors expect the company to finish the year with a run rate between $100 billion and $120 billion and to seek a public valuation above $2 trillion; its late-May round valued it at $965 billion. The next day, Bloomberg reported that the pre-IPO revolving credit facility is set to exceed its roughly $10 billion target, with the most active arrangers asked to commit about $1.25 billion each, a second tier around $1 billion, and less active roles at about $750 million or less. Talks are ongoing and the company could hold the revolver at or below target.

For comparison, OpenAI's run rate was reported at $40 billion on August 13, double its $20 billion mark at the end of last year. The two may calculate the metric differently and the comparison has limits, but the growth gap is striking: Anthropic went from $9 billion to $65 billion in seven months. These two figures should be read against the procurement share data above — Anthropic leads by roughly four points in the Ramp sample, while OpenAI is growing faster quarter-to-date. Total revenue and enterprise customer count are not the same thing; Anthropic's revenue mix skews toward high-priced coding and agent workloads with heavy per-customer consumption, which explains why its share lead is narrower than its revenue lead.

The credit facility says more than the revenue figure does. When a company arranges a ten-billion-dollar revolver before listing, the money covers the timing mismatch between compute commitments and build-out spend rather than funding day-to-day operations. That tells you compute purchases for the next one to two years are already scheduled, and that after listing those commitments will sit on quarterly statements as fixed costs. Patient private capital can tolerate pricing below cost for a long time in exchange for penetration; the moment quarterly disclosure begins, gross margin becomes something the company is publicly questioned about. When signing multi-year commitment contracts, how long price-protection clauses run and whether the vendor can unilaterally swap model versions and deprecation schedules will matter more and more as the listing approaches.

Two opposite roads for inference silicon

On August 17, Groq announced a $350 million Series A at a $3.5 billion valuation, led by Disruptive with planned participation from NVIDIA, bringing recent funding to $1 billion alongside $650 million raised in June. The line in the announcement that deserves most attention is the identity description: Groq is an NVIDIA Cloud Partner, certified to design, deploy, and operate NVIDIA accelerated computing to NVIDIA's reference architecture and operational standards, and the capital will support customers seeking medium and larger NVIDIA clusters for training and inference. The company runs 13 data centers, serves roughly 6 million developers, and plans to scale from 54 megawatts to more than 200 megawatts in 2027. Executive Chairman Alex Davis also cited the team's experience operating LPUs at scale, so the in-house silicon track has not ended. The transaction remains subject to customary closing conditions.

A company that started with its own inference silicon and was long treated as the alternative to NVIDIA taking NVIDIA's money and building to NVIDIA's architecture says that, in this window, selling capacity turns into cash faster than selling a chip architecture. The reasoning is not hard to infer: incremental inference demand is concentrated in workloads already built on CUDA, capturing it with a proprietary architecture requires a complete software stack, and capacity can be rented out immediately.

A signal pointing the other way arrived the same week. On August 18, Etched closed a $700 million Series D at a $21 billion valuation led by Jane Street, roughly one month after a $10.3 billion Series C on July 23 — a doubling. Jane Street reportedly tested the chip before investing. Etched's Sohu is an ASIC that runs Transformer models only, fabricated on TSMC's 4nm process; the cost is that it cannot run other model types, and the return is better energy efficiency at the same node. A valuation that doubled in a month rests on a single test run by the investor, with no independently published third-party benchmark.

Put the two signals side by side and what dedicated inference silicon lacks right now is not money but consensus. Those betting on dedicated architectures are betting the Transformer shape is stable enough to be worth fixing into silicon; those turning to an NVIDIA cloud are betting that what enterprises lack most right now is capacity that can be delivered immediately. In procurement terms, the next two to three years will bring more inference suppliers to choose from, and whether you can switch between them is settled at the software layer, not the hardware layer — whether the same workload can move between NVIDIA, dedicated silicon, and domestic Chinese accelerators decides how much of this new supply can actually serve as negotiating leverage.

China: token calls up ninefold, advertising down a fifth

On August 18, Baidu reported second-quarter results: total revenue of RMB 31.33 billion, down 4% year over year and below expectations, because online marketing revenue fell 19% to RMB 13.1 billion and that decline outpaced growth in the AI business. The AI figures look entirely different: AI cloud infrastructure revenue grew 50% year over year to RMB 7.3 billion, GPU cloud revenue within it grew 283%, and external customer token usage on the Qianfan platform grew more than ninefold. Core AI-powered business revenue — cloud infrastructure, AI applications, and AI-native marketing services — came to RMB 12.5 billion, up 25%, roughly half of overall business revenue.

That ninefold figure says more than the 50% revenue growth does. The gap between token usage growth and cloud infrastructure revenue growth means unit token prices fell sharply over the same period, with most of the demand growth absorbed by price declines. That matches how China's model market has priced over the past two quarters, and it explains why Chinese cloud vendors' AI revenue growth looks less dramatic than their usage growth. The 283% GPU cloud figure points to something else: enterprises are starting to rent compute and run models themselves rather than only calling APIs. Put the two numbers together and the procurement signal is quite specific: Chinese enterprise AI consumption is scaling fast while budget growth lags far behind, and the cost pressure has been pushed onto the supply side. Multinationals with China operations have one more layer to account for: the unit-cost gap between the two markets for the same workload keeps widening, while compliance requirements, data residency rules, and available model lists differ on each side, which raises the case for setting routing and quotas separately by region.

Compute supply: short contracts push annual value to $40 million per megawatt

The most substantive change upstream over the past week was in contract structure rather than headline prices. According to The Information on August 16, Nebius and CoreWeave are shifting sales toward short-term compute contracts. Nebius now cites roughly $40 million in annual contract value per megawatt of data center power, with short-term deals ranging from $40 million to $50 million per megawatt and the first contract already signed; market expectations for that metric a year ago were $10–15 million. AWS over the same period continues to favor long-term deals.

The choice of contract length reveals what suppliers expect. Long contracts lock in revenue and smooth volatility, which is what supply-rich conditions produce; short contracts keep pricing power in the seller's hands, which is what suppliers do when they expect prices to keep rising. Two figures support that read: prior-generation GPU prices are up more than 30% versus January through March, and CoreWeave's near-term capacity remains effectively sold out after raising prices across its lineup by roughly 25% in July. One of the ready-made ways enterprises have been holding costs down is narrowing as a result. For the past few quarters, using the price gap between neoclouds and the three major public clouds to lower inference cost worked well; that elasticity is now shrinking — the gap persists, but the neocloud side is also raising prices and is no longer willing to lock low prices to customers under long contracts.

Memory prices show no sign of turning. TrendForce's July 9 forecast put third-quarter server DRAM (server memory; every accelerator card in an AI server needs large amounts of it attached) contract prices up 13% to 18% quarter over quarter, with the narrower increase attributed to long-term agreements capping price rises and to weaker consumer demand. Spot prices are still setting records: on August 7, DDR4 1Gx8 3200 spot reached $42.45, the highest on record for that measure. On the server side, 64GB RDIMM contract prices have climbed from $450 in the fourth quarter of last year to above $900 in the first quarter. Customers with long-term agreements and those without carry entirely different costs through this cycle, and enterprises building their own inference clusters will feel the pressure sooner than those renting.

There is one concrete figure on the demand side: Baidu's GPU cloud revenue grew 283% year over year. It corroborates the neoclouds' price increases — demand from enterprises renting compute to run models themselves is scaling fast, and suppliers have already adjusted their quotes accordingly.

The line drawn this week

The dividing line between model vendors and enterprises moved onto data this past week. OpenAI leaves the context in the customer's hands and takes back only alert signals; Anthropic holds the content itself for thirty days and publicly admits this will hurt the business. The two diagnose the problem identically and choose opposite commercial answers, and Ramp's procurement data indicates enterprises have already begun voting along that line. This continues the direction of the moves two months ago: vendors are unbundling control and pricing the pieces separately, giving away the most replaceable part and keeping the leverage. What differs this time is that the part being unbundled is the data itself, and data custody is frequently decided by external regulatory obligation, leaving enterprises little room to compromise.

There is a new item on the cost sheet. OpenAI put a price on the compute that safety monitoring consumes: a fifth of the inference compute being monitored. That cost will not disappear; it enters the price list in some form. Over the same period, frontier delivery cadence became unpredictable under safety review — the largest planned frontier run is still paused, with no timetable given. Treating an unreleased model tier as a fixed item in next year's roadmap carries considerably more risk than it did a year ago.

Money keeps pouring into this supply chain while the side selling compute tightens its terms. Anthropic is arranging a ten-billion-dollar revolver and preparing to list above $2 trillion, Stripe paid more than $7 billion for the model aggregation layer, Etched doubled its valuation in a month, and Groq turned to building an NVIDIA cloud — capital's pricing of this chain is still rising, and the next two to three years will bring more compute and model supply to choose from. At the same time, the neoclouds are shortening contract terms and raising quotes, taking back part of the certainty around falling prices. What enterprises can count on is more choice; what they cannot count on is which vendor's quote stays cheap and which vendor's policy stays stable. Under that structure, hold access rules, routing policy, cost attribution, and the audit chain in your own hands, and however vendor forms, price lists, and data policies shift, none of it reaches your own rules or your own books.