Seven domestic AI chip companies that are listed or preparing to go public generated approximately RMB 21.5 billion in revenue in the first half of the year. Among the six that disclosed inventory levels, their balance sheets already show RMB 23.9 billion in inventory. Of the five that disclosed operating cash flow, their combined cash outflow over six months totaled about RMB 6.4 billion—only Cambricon remained positive. This RMB 23.9 billion should not be simplistically interpreted as “chips that can’t be sold.” Cambricon is a good example: of its roughly RMB 9 billion inventory value, RMB 5.7 billion consists of raw materials, RMB 2.5 billion is tied up in manufacturing, and only about RMB 700 million represents finished goods or products already shipped and awaiting customer acceptance. Money has already been paid months in advance to wafer foundries, memory suppliers, substrate manufacturers, and packaging providers—but revenue can only be recognized and collected after the chips are fully assembled, servers are configured, and customers complete acceptance. The faster orders grow, the more working capital is required upfront. As a result, several domestic GPU manufacturers are now seeking external parties to absorb this cash cycle burden. Cambricon’s inventory rose from RMB 4.9 billion to RMB 8.2 billion in six months; advance payments to suppliers surged from RMB 700 million to RMB 2.9 billion; meanwhile, its accounts payable bills skyrocketed from RMB 350 million to RMB 3.09 billion, indicating a heavy reliance on bank-accepted drafts for procurement. Moore Threads is even more extreme: with RMB 1.74 billion in revenue over six months, its inventory reached RMB 3.55 billion, operating cash flow net outflow hit RMB 2.17 billion, and long-term borrowings climbed to RMB 2.83 billion. Breaking down the seven companies, their inventory is ultimately supported by four sources of funding: Extended payment terms from suppliers and banks, advance payments from customers, bank loans, and continuous inflows of new equity capital. This is where I find the most cause for concern. Roughly calculating based on annualized cost of goods sold relative to inventory levels, Cambricon carries roughly 445 days of inventory, Moore Threads about 590 days, while NVIDIA only manages around 108 days. While the exact calculation methods aren’t identical, the order-of-magnitude difference is enormous. Moreover, NVIDIA is now doing something fundamentally different: it has the capacity to offer its largest customers payment terms of nearly one year—effectively using its own balance sheet to finance its customers. The cash flow pattern among domestic AI chipmakers is precisely the opposite: suppliers, banks, and shareholders are still financing the chip companies, which then use that capital to fund their own inventories. Therefore, as China’s domestic substitution enters its second phase, I’m less interested in how many more TOPS have been achieved or which benchmarks have been surpassed. The next performance metrics that truly matter are: inventory turnover, accounts receivable, and operating cash flow. Building a chip only proves technical feasibility; securing orders only proves customers are willing to try. Only when customer payments can fund the next batch of wafers will China’s AI chip industry truly become a self-sustaining business.