#股票交易分享挑战

Open-sourcing equals congestion for Kimi—Is the $50B valuation of the Darker Side of the Moon a bubble?
On one side, after open-sourcing, compute clusters were “crushed” by a surge of requests; emergency closure of subscription channels for new users followed. On the other side, capital rushed in, and the valuation soared to $50B. The domestic AI unicorn, the Darker Side of the Moon, in the midsummer of 2026, is undergoing a double tempering of ice and fire.
Late night on July 27, the Darker Side of the Moon announced the open-sourcing of the Kimi K3 model weights, while also releasing a technical report and three core underlying infrastructure technologies. This mixture-of-experts model with a parameter scale of 2.8 trillion became, in one move, the open-source large model with the largest parameter count in the world. On the Frontend Code Arena programming leaderboard, it scored 1679 points to surpass Claude Fable 5 and GPT-5.6 Sol, taking first place on the chart. After the open-sourcing news broke, the global developer community instantly boiled over. The CEO of Hugging Face publicly stated that within just 30 minutes of the K3 launch, it received over 4,000 likes.
However, the other side of the technological celebration is the severe pain of infrastructure. A blowout in technical demand coupled with insufficient supply capacity created a glaring mismatch this summer. But the capital market reinterpreted this “pain” in exactly the opposite way—as the biggest positive.
I. The K3 cluster being “squeezed out”
On July 19, the Kimi team of the Darker Side of the Moon issued an announcement: due to an unexpected challenge in compute capacity, effective immediately it would pause subscriptions for new C-end users, putting all existing compute capacity into serving already-subscribed users, while pushing compute expansion at full speed.
Behind this announcement is a real “compute squeeze.” After K3 was open-sourced, the volume of user requests grew exponentially. The existing compute architecture could not handle the load. Service queues and response latency followed, and ultimately the team had to hit the pause button for new user subscriptions. The Darker Side of the Moon said bluntly that K3’s overall performance still lagged the strongest closed-source model, but the traffic flood brought by open-sourcing clearly exceeded the team’s expectations.
The compute architecture design did not sufficiently anticipate the traffic flood, and there are shortcomings in elastic scaling. Even more troublesome is that procurement cycles for scarce resources like GPUs are long and deployment costs are high. Temporary capacity expansion faces the real-world problem of hardware supply shortages. According to industry data, in 2026 the supply-demand gap for domestic AI servers is on the order of 180k to 390k units. Meanwhile, among the factories in China that are the only ones capable of volume-producing 7nm-class chips, Semiconductor Manufacturing International Corporation (SMIC) has monthly capacity of fewer than 20k wafer starts—this number is only one-fifteenth of TSMC’s capacity for the same node. If the team urgently calls in third-party cloud resources or leases idle compute, the unit cost could be several times higher than using its own cluster, further squeezing profit margins.
Kimi is not an isolated case. Multiple AI companies face the dilemma of “open-sourcing equals congestion,” which exposes the fragility of AI infrastructure as a whole. The issue of compute structural misallocation is prominent: older general-purpose compute is relatively over-supplied, while advanced compute resources suited for training massive models—featuring high-speed interconnects and mature software ecosystems—remain persistently tight. Compute that is nominally deployed in a data center is not necessarily effective production power that can be efficiently called upon.
II. The tighter the compute, the crazier the capital
Interestingly, the tighter compute gets, the more excited capital becomes.
On the very same day that Kimi announced it would pause subscriptions for new users, market rumors spread: the Darker Side of the Moon had completed a valuation-round financing of about $31.5 billion during the summer, and planned to start Pre-IPO financing negotiations in August, targeting a pre-investment valuation as high as $28k and the fastest possible listing on the Hong Kong Stock Exchange within the next 6 months. Longtime shareholders such as Alibaba, Tencent, and Sequoia kept adding more. Founded only three years ago, the company’s cumulative financing has already exceeded 37.6 billion RMB.
When compute is “crushed,” capital views it as “demand validation.” Users’ real needs are strong, and market penetration far exceeds expectations—these are the growth signals capital values most. Data shows that in the first quarter of 2026, the Darker Side of the Moon’s annual recurring revenue (ARR) surpassed $100 million; in May it surpassed $200 million; in June it further surpassed $300 million, achieving a threefold jump within three months. Revenue growth mainly comes from API calls, enterprise-level services, and overseas market expansion, with API accounting for about 70%.
Compute itself has become the “oil” of a new track. Whoever controls more compute has the voice. Capital interprets compute tightness as a signal that “the moat is deepening,” while the company, following the trend, packages “acute shortages of compute” as a financing narrative of “needing more capital to build the next-generation infrastructure,” driving valuations higher. In the frenzy, capital temporarily ignores near-term profitability issues and focuses instead on market share and user base—burning cash to gain scale is considered the best strategy right now.
A research report from Citic Securities points out that Kimi K3 shows the Darker Side of the Moon’s ongoing technical accumulation and competitiveness. Under constrained compute capacity, Chinese model companies keep iterating based on technical taste and organizational efficiency, working to narrow the performance gap with the top American models. But the report also reminds that long-term sustainability still needs continuous validation of progress in commercialization.
III. Double scrutiny: the path to profitability and the timeline
Amid the capital celebration, a more fundamental question gradually emerges: Can commercial sustainability truly hold up?
On the revenue side, although there are many users, will the paid conversion rate and average revenue per user keep up? Can a free open-source model support long-term operations? While ARR is climbing quickly, there is a huge gap between $300 million in annual revenue and a $180k valuation.
On the cost side, a black hole is widening. Compute procurement and maintenance form the biggest fixed-cost items and rise non-linearly as users grow. R&D investment, talent competition, and marketing expenses stack up simultaneously, keeping total spending high. Industry analysis suggests that in 2026, over 85% of domestic AI chip capacity depends on SMIC. But the production bottleneck for 7nm-class chips is unlikely to be broken in the short term, meaning compute costs will remain high in the foreseeable future.
The exploration of a profitability path is already underway. In the short term, differentiated services to enable tiered pricing—such as priority usage and customized deployments—may be the most direct monetization approach. In the medium term, B-side monetization such as enterprise-level APIs, industry solutions, and model fine-tuning services is accelerating in implementation. The Darker Side of the Moon has already launched the Kimi Hosted Agent platform for enterprise customers, helping companies embed AI capabilities into business workflows like office work, R&D, and investment research through standardized interfaces. In the long term, reducing unit costs through technical breakthroughs is the real solution. The K3’s use of KDA linear attention mechanisms has improved scaling efficiency by 2.5 times, but it still has not reached the breakeven point.
Uncertainty in the timeline is the biggest worry. Under the current burn rate, if the company cannot reach breakeven within 2 to 3 years, it may face risks such as a financing gap or being forced into strategic contraction. Once it lists, quarterly performance targets in capital markets will become even more stringent, and profitability pressure will only increase. $KIMI