A report titled "2026 State of Enterprise AI" released by Menlo Ventures, a well-known Silicon Valley venture capital firm, shows that of U.S. enterprise spending on open-source models, 89% went to Chinese models. Among the top three, Kimi accounted for 61%, DeepSeek 15%, and GLM 11%.

The market is quite willing to embrace a simple narrative: on the strength of performance and pricing, domestic large models are capturing overseas developer budgets.

But the same report contains another set of figures: although open-source models have already captured nearly one-fifth (19%) of Token usage, they actually account for only 4% of total enterprise model spending, with the vast majority of enterprise AI budgets still flowing to closed-source large models.

In other words, 89% represents a highly concentrated but limited-volume share.

Volume itself is not the biggest issue; the real change lies in the cost structure. You see, selling tokens looks like selling compute, but in essence it is delivering state. And state does not attach to compute—it resides in memory and storage, which are precisely the components where this round of price-increase pressure is most concentrated.

The contradiction thus emerges: revenue is accounted by Token, but costs follow GB.

01. Traffic Is Running, Assets Are Sitting Idle

As the market closely watches the continuously updated Token usage rankings on major aggregation platforms and hotly debates the share battle of domestic large models in overseas open-source markets, most analyses still anchor the narrative on model capabilities and pricing advantages, assuming this is ultimately a "compute-for-Token" business.

This understanding holds up in traditional dialogue scenarios, but in the Agent tasks that currently drive overseas incremental traffic, the underlying logic of workloads is beginning to diverge.

The resource model corresponding to traditional dialogue inference is intuitive enough and is also the foundation supporting most of today's API pricing systems.

In this ideal framework: a user initiates a query, the inference cluster completes computation and returns content; once a round of dialogue ends and the session closes, the cache resources occupied by this request are released, and hardware returns to idle.

Service providers can estimate costs based on total Token volume. When cluster utilization still has headroom, higher throughput offers more opportunity to dilute unit overhead. The logic of low-price volume and trading scale for profit is thus self-consistent, and Token count and resource consumption are assumed to be on similar growth trajectories.

But the main traffic force driving incremental growth in overseas aggregation channels is shifting from scattered human-machine dialogues to Agent tasks with automated pipelines.

The resource rhythm of such workloads diverges significantly from ordinary Q&A: at task startup, some scenarios require injecting hundreds of thousands of Tokens worth of prerequisite documents and batch business data; then comes dozens of rounds of chained tool calls, continuously retrieving and reading external information; the entire session may hang for hours or longer, with the same prerequisite context being repeatedly reused.

More critically, without actively closing the session, this context cache will continuously occupy resources. The model keeps generating new Tokens, but appending to cache does not automatically release this occupancy back to the resource pool.

As the proportion of long-session Agent traffic rises, the constraint point is also shifting: it is no longer just about inference speed—the upper limit of concurrent resident sessions the system can support becomes the new binding constraint.

The million-level context windows vendors advertise externally are more of a product selling point for customers; on the resource ledger of the data center, it means that once a session uses up the window, each active concurrent session may need to reserve tens of GB of HBM space to carry the continuously online session state.

KV quantization, MLA, and various cache compression technologies on the market are all trying to reduce per-session VRAM overhead and raise the concurrent scale a cluster can support.

But they are essentially only optimizing unit costs and can hardly bypass that underlying constraint: as long as a session is in an active state ready to respond at any time, its core state data needs to reside in high-speed storage.

The core increment of this round of Token traffic expansion does not entirely come from newly generated output text.

Trillion-level call volumes easily create an illusion: growth comes from an endless stream of new content generation. In reality, a considerable portion of traffic growth corresponds to dynamic state assets that continuously stack up in the cluster, remain locked for extended periods, and are difficult to reclaim for cross-task scheduling at low cost.

Platform statistics panels often only record the number of Tokens in and out, but in the service provider's operating ledger, the truly costs that continuously accumulate come from those VRAM and high-speed memory resources that are occupied long-term and difficult to reallocate to other businesses at low cost.

The contradiction thus emerges: the industry's mainstream pricing framework is still derived from the logic of the instantaneous compute era, with revenue measured by input and output tokens. But the corresponding capital expenditure and high-speed VRAM occupancy costs accumulate continuously with the residency time of concurrent sessions.

Some Agents may hang in operation for a whole day, producing very few Tokens per unit time with limited revenue, yet VRAM remains continuously occupied. It inflates the call volume on the rankings, but revenue cannot keep up with the continuous resource consumption—the two growth curves that once tended to synchronize thus open up a gap.

This misalignment between cost and billing may be slowly rewriting the competitive background color of the Token overseas expansion track.

Previously, the competitive path in this track was simple and direct: compress per-Token compute overhead, acquire volume with low prices, and dilute hardware investment through scale. When the workload structure tilts toward Agents, clinging to the single competitive dimension of throughput and pricing may hit the trap of rising traffic but narrowing profits.

Rising call numbers easily send a signal of business improvement, but some service providers tend to overlook that accumulating session state assets are posing a continuous erosion risk to gross margins.

Beyond the bustling share rankings, the management capability of state assets may be forming a new invisible dividing line.

02. Token Overseas Expansion, the Invisible VRAM Liability

When storage costs rise, the market easily simplifies it into a story of storage chip price increases. But viewed within the framework of inference cluster operations and capital expenditure, component spot prices are just one of many variables.

What constitutes a sustained rigid constraint on profits is often a whole chain of rules interwoven from tied supply, asset depreciation, cross-border traffic, and customer compliance.

Upstream storage chips are in a structurally tight balance cycle. Long-term agreement quotes for server DRAM, enterprise-grade NAND for AI clusters, and HBM continue to rise, with original manufacturers' overall inventory maintained in a low range.

To secure stable supply, service providers building their own inference clusters can hardly rely solely on spot purchases for continuous expansion; they often need to sign quarterly or even semi-annual minimum-purchase long-term agreements in advance.

This also reshapes the nature of the investment: hardware procurement is no longer a variable input that can be flexibly adjusted with business rhythm, but has become capital expenditure with rigid performance obligations.

Once such contracts are finalized, even if subsequent overseas traffic cannot fully fill cluster capacity, contract payments and asset depreciation must still be honored and accrued on schedule, and revenue fluctuations can hardly offset this rigid expenditure.

There is another cost that is easily masked, coming from compliance retention requirements of overseas enterprise customers. Many overseas enterprises, when purchasing inference services, require complete audit logs, call chains, and even original session content to be traceable and auditable long-term.

This persistent storage is independent of the KV cache temporarily generated during inference and will occupy storage capacity stably over the long term.

Most standardized Token procurement contracts do not separately itemize this storage service fee; the corresponding overhead is spread into the unified Token quote, making it difficult to separate and identify externally.

Even with KV cache compression, quantization schemes, and iterative model versions for optimization, while the physical service life of cluster hardware remains, the effective business lifecycle adapted to current Agent workloads is being continuously compressed.

This forces service providers to lower the expected return period of assets, and the depreciation allocation per unit Token rises accordingly. With multiple constraints stacking, storage is no longer just a supporting cost of servers—under such workloads, it is approaching the weight of compute and often becomes the bottleneck of cluster expansion.

Storage's erosion of profits is not a simple, general cost increase, but a continuous suppression of cluster asset returns along multiple independent paths. Such signals are not broken down in external quotes; most can only be observed in the service provider's back-end operations and business reports.

First is the prefix cache hit rate trap. The underlying logic of prefix caching is trading storage for compute reuse: caching frequently repeated context prefixes so that once hit, repeated computation can be skipped, reducing per-Token compute overhead.

But the cache itself carries write overhead, eviction scheduling logic, and lifecycle constraints.

Once traffic becomes highly fragmented, tenants scattered, and request randomness increases, the cache hit rate may quickly fall below the break-even point. Large amounts of VRAM reserved for cache continuously occupy capital but can hardly deliver the expected compute savings, thereby dragging down the asset return rate of the entire cluster.

Second is the idle depreciation brought by peak reserve. To handle burst traffic, avoid forced session eviction, and prevent service stability degradation, operations teams often need to reserve sufficient storage and VRAM redundancy.

Redundant resources are reserved for spikes and mostly sit idle in normal times. Redundancy improves stability, at the cost of lower asset turnover efficiency. Under cross-border tidal traffic, this depreciation cost remains rigid and difficult to fall in sync with off-season Token revenue.

Finally, there is the silent resource lock-up caused by long sessions. Some Agent automation tasks have this characteristic: low Token output rate per session, but extremely long session lifecycle, continuously occupying a segment of VRAM space.

Such workloads do not quickly inflate call volume on platform throughput rankings, but they gradually squeeze VRAM resources, continuously consume cluster expansion headroom, and may force service providers to initiate a new round of hardware procurement ahead of schedule.

The market is accustomed to judging the competitive landscape of domestic large-model overseas expansion by Token unit price and model capability, but under the new traffic structure dominated by Agents, the competitive focus is extending from pure compute competition to a more refined asset efficiency management.

While industry attention remains on Token call rankings, storage asset turnover, cache break-even lines, and depreciation rhythm are slowly becoming hidden yardsticks affecting the profitability of overseas business.

03. Will Storage Inflation Reshape the Large-Model Overseas Business?

The upward pressure of storage costs draws a clear dividing line in the distribution chain: players holding hardware assets that carry long-resident sessions will bear the most rigid pressure that is hardest to pass on externally.

Among them, some service providers control scheduling strategies and cache rules and have the initiative in downstream pricing, but their corresponding risk exposure is also highest.

Many suppliers building their own clusters need to rely on quarterly or semi-annual minimum-purchase long-term agreements for stable supply, and are likely to bear alone the depreciation pressure from rising chip quotes, asset idleness caused by cross-time-zone traffic mismatch, and rapid hardware iteration.

During the stage of tight storage balance and continuous expansion, the incremental revenue brought by traffic growth is easily devoured quickly by continuously added capital expenditure.

Another mainstream approach is to reach revenue-sharing hosting partnerships with overseas cloud vendors. From the perspective of the industry ledger, this is often understood as a colloquial off-balance-sheet treatment of assets and liabilities: the model side outputs model authorization, while the hardware investment and asset risks corresponding to memory, persistent storage, cross-time-zone operations, and compliance log retention are mainly borne by the cloud vendor.

But this is not equivalent to asset divestiture in the accounting sense, nor is it equity investment. If the contract sets minimum usage or cost-linked clauses, part of the cost pressure may still be transmitted back.

In addition, there is a path of only outputting model weights and relying entirely on aggregation platforms for distribution. Such suppliers basically do not hold inference hardware and have the smallest direct exposure to storage and compute assets, but their bargaining power is usually the weakest.

Platforms control routing allocation, traffic quotas, and cache strategy adjustment permissions, while suppliers can mostly only passively accept the traffic allocated to them. Even if load characteristics can be fine-tuned through versions and parameter configurations, it remains difficult to autonomously intervene in the asset efficiency of the underlying cluster.

There is no absolutely perfect path; these choices are mostly trade-offs between traffic growth potential and storage asset exposure, while also balancing other constraints such as compliance, customer ownership, and pricing initiative.

In a storage price upcycle, asset off-balance-sheet treatment will become a more attractive choice for many players; only when upstream capacity is released and long-term agreement quotes fall will the profit window for self-built clusters reopen.

The endless stream of KV quantization, sparse attention, CXL memory pooling, and tiered caching solutions in the industry easily creates an illusion: that technical optimization alone can erase storage cost constraints.

But on the operating ledger, all engineering solutions have clear ceilings. They can raise asset utilization but cannot eradicate the rigid storage demand on the state side.

KV compression can reduce per-session VRAM occupancy, often at the cost of extra compute overhead; sparse attention can cut computation but cannot eliminate the residency requirement of session state; CXL memory pooling helps offload VRAM pressure, but server hardware modification itself brings a new round of capital investment.

More critically, optimization returns are highly tied to traffic structure: Agent tasks with high context reuse can fully enjoy cache benefits, while returns on optimization for fragmented and one-off requests decay rapidly.

Technical means can raise the ceiling of asset efficiency but can hardly bypass the underlying constraint that session state must rely on hardware storage.

In the long run, storage-compute separation architecture has the opportunity to deeply decouple computation from state storage, enable flexible migration of session state across nodes, and reshape the asset allocation logic of clusters.

But the reconstruction chain of the entire infrastructure is long and can hardly reverse the current cost landscape in the short term.

Compared with call volume rankings that only reflect scale, another set of metrics better mirrors the true quality of this business: long-context request proportion, cache hit rate, average session survival time, overseas node storage idle rate, and quarterly storage long-term agreement quotes.

A simple test yardstick is: in the process of continuously rising call volume, can per-unit Token gross margin be maintained or even improved?

If revenue continues to rise but gross margin stays flat or even declines over the long term, it often signals that incremental gains are being continuously eroded by storage depreciation.

Token overseas expansion is undergoing a quiet business format shift: early on it was more easily classified as a compute rental business, with marginal costs expected to be continuously diluted with scale. But as long-session Agents gradually become the main traffic force, the nature of this business is moving toward online state hosting.

GPU memory and storage have become core accumulated assets, and marginal costs are unlikely to continue declining; asset turnover directly determines the profit foundation.

Traffic can migrate at any time along the routing of aggregation platforms, but once hardware capital expenditure is deployed, the depreciation schedule will not adjust in sync with customer traffic. Matching highly liquid short-term orders with heavy assets that are depreciated over the long term naturally creates a cyclical mismatch between revenue and costs.

The market is always comparing total token call volume and share rankings, but from the supply-side ledger, the more worthwhile question is: how many GB of storage must be locked in for each additional persistent session, and can this asset recover enough cash flow within the depreciation cycle?

Share is the vote cast by the demand side, while the storage asset ledger is the core constraint that determines whether the large-model overseas business can run sustainably over the long term.

As tokens gradually move toward standardized commodities, what ultimately creates the profit gap may no longer be the compute used to generate text, but the storage assets that carry context memory.