LLM Funding and Post-IPO Declines: Navigating Commercial Sustainability Under Compute Cost Pressure

1. The Paradox: Hot Private Markets, Cold Public Markets

The LLM sector has recently entered a notable contradictory state. On one hand, foundation model startups are still closing large-scale funding rounds with climbing valuations. On the other hand, several companies that have gone public have experienced significant share price declines and market cap erosion. Some model-focused companies listed on the Hong Kong Stock Exchange have fallen below their IPO price, while related listings on both A-share and Hong Kong markets have undergone sharp pullbacks. Meanwhile, leading AI labs continue to negotiate new financing rounds, and the industry’s overall capital appetite has not diminished simply because some players have gone public.

This “hot private, cold public” dynamic reflects a fundamental split in how capital markets assess LLM companies. Private investors are still betting on generational technology leaps and winner-take-most concentration effects. Public markets, however, are increasingly applying traditional financial scrutiny—examining earnings trajectory, revenue quality, and the pace of cash burn.

2. Compute Cost: The Core Cash Outflow for LLM Companies

Understanding why LLM companies remain perpetually capital-hungry requires a close look at their cost structure.

2.1 Training-Phase Compute Consumption

Large-scale pre-training demands thousands to tens of thousands of accelerator cards running continuously for weeks or months. Based on widely understood industry benchmarks, a single full training run of a model at the hundred-billion-parameter scale can consume tens of millions to hundreds of millions of dollars in compute alone. More critically, training is not a one-time event. Model iteration, multimodal expansion, long-context support, and reinforcement learning alignment all require repeated large-scale compute cycles.

This means that even after completing one generation of a model, development of the next generation is already underway. Training cost is not a fixed, one-off capital expenditure—it is a continuously rolling operating expense.

2.2 Inference-Phase Elastic Costs

Compared to training, inference costs are more subtle but equally significant. As token consumption grows rapidly, the compute required on the inference side scales linearly or even super-linearly. Publicly reported data suggests that some model companies have seen token volume grow by dozens of times within a single half-year. This pace places extreme demands on inference cluster scaling, scheduling efficiency, and cost control.

Key variables in inference cost include:

• Concurrency scale: The number of simultaneous requests determines the minimum compute pool required.
• Sequence length: Long-context requests substantially increase per-inference KV cache and computation.
• Output length: Generative tasks typically produce far more output tokens than input, directly affecting cost.
• Model size and precision: Larger parameter counts and higher numerical precision both drive up per-token cost.

2.3 Hidden Engineering Costs

Beyond raw compute, data engineering, model evaluation, safety alignment, and online service reliability all demand substantial engineering resources. These investments do not appear directly on a compute bill, but they significantly extend development timelines and raise personnel costs.

3. Revenue Growth and Widening Losses Can Coexist

An easily overlooked reality is that rapid revenue growth does not guarantee shrinking losses. An LLM company can see revenue multiply several times year-over-year while its losses expand simultaneously. The reason is straightforward: when the growth rate of compute costs exceeds the growth rate of revenue, scaling up actually accelerates cash consumption.

Key metrics for evaluating this dynamic include:

• Unit economics: Whether revenue per million tokens covers the corresponding inference cost.
• Gross margin trajectory: Whether inference costs show marginal reduction as scale increases.
• Customer mix: The proportion and payment stability of enterprise versus individual users.
• Geographic revenue split: Pricing and willingness-to-pay vary significantly across markets.

For example, some model companies have reported annual revenue in the hundreds of millions of yuan with year-over-year growth exceeding 150%, but with a high share of overseas revenue. This exposes their revenue structure to exchange rate fluctuations, cross-border compliance requirements, and regional competition. When growth is driven primarily by rising token consumption—and token pricing cannot fully offset inference costs—expansion becomes a source of cash flow pressure rather than relief.

4. Why Continuous Funding Remains Necessary

Given the cost and revenue dynamics above, the logic behind ongoing fundraising becomes clear.

4.1 The Technology Race Allows No Pause

LLMs are still in a phase of rapid evolution. Multimodal capabilities, reasoning, tool use, and AI agents are all advancing in parallel. Any company that slows its compute investment will see its model capabilities fall behind quickly. Companies must therefore maintain ample capital reserves to ensure their R&D cadence is not constrained by cash flow.

4.2 Scaling Inference Requires Upfront Investment

Building and optimizing inference clusters demands significant upfront capital expenditure. Even if per-unit costs can eventually decline through optimization, companies must first deploy sufficient compute capacity to meet rapidly growing demand. This “invest first, monetize later” pattern further widens near-term funding gaps.

4.3 Commercialization Is Still in Early Validation

The commercialization playbook for LLMs is not yet fully written. Subscription models, API usage, enterprise solutions, and agent distribution are all being explored simultaneously. Until business models are thoroughly validated, capital remains the essential resource for navigating uncertainty.

5. An Engineering Perspective: How to Gauge Proximity to Profitability

For technical practitioners, assessing the commercial sustainability of an LLM company can be approached through several dimensions where engineering and finance intersect.

5.1 Rate of Inference Cost Convergence

Watch the trajectory of per-token cost reduction. Common optimization techniques include model quantization, sparsity, distillation, speculative decoding, batch optimization, and inference cluster scheduling improvements. If per-unit cost continues to decline with scale, it signals that engineering capability is translating into economic advantage.

5.2 Share of Non-Token Revenue

Revenue driven purely by token consumption is vulnerable to price competition. Solutions built on top of model capabilities—private deployments, value-added services, and tailored enterprise offerings—tend to offer better margin profiles. Tracking the trend in non-token revenue as a share of total revenue provides a clearer picture of commercialization quality.

5.3 Customer Retention and Usage Depth

Whether enterprise clients are steadily increasing their call volume is a strong indicator of product value. Token consumption growth driven by deeper usage from existing customers suggests the model is creating real business stickiness. If growth is mainly from one-time trials by new customers, it warrants caution.

5.4 Capital Efficiency: Capex to Revenue Conversion

Examine how much incremental revenue each unit of new capital expenditure generates. This ratio reflects capital efficiency. In a high-compute-cost environment, differences in capital efficiency will directly determine which companies can go further on the same funding base.

6. A Sober View of the Gap Between “Cash Burn” and “Profitability”

The capital demands of the LLM industry are real and ongoing—not something that can be dismissed as simple hype. The rigidity of compute costs, the speed of technical iteration, and the uncertainty of commercialization collectively mean this sector will require substantial capital injection for a considerable period.

At the same time, public market price signals are sending a clear message: the tolerance for perpetual cash burn is finite. Once a company goes public, its financials face far more rigorous scrutiny. Whether unit economics improve, gross margins turn positive, and cash burn rates decline within a reasonable timeframe will become the defining metrics of corporate value.

For technical decision-makers, what truly matters is not the headline figure of any single funding round, but whether that capital can be effectively converted into generational model improvements, sustained per-token cost reductions, and scalable commercial revenue. Only when these three elements form a virtuous cycle can an LLM company genuinely reach the point where continuous external funding is no longer necessary.