BlockBeats News, July 7th - Six recent research reports from institutions including top Wall Street investment banks outlined a new AI research finding: NVIDIA is transitioning from a chip supplier to a compute financing system's credit fulcrum; TSMC is accelerating N3, N2, and advanced packaging capacity expansion to match AI demand; Samsung Electronics and SK Hynix's memory cycle remains strong, but short-term faces a cooldown in crowded trades; Japanese material companies such as MLCC release film manufacturers, and optoelectronic module vendors including Nichia and Epistar are still favored.

SemiAnalysis

A deep-dive research by SemiAnalysis takes the issue a step further: AI infrastructure is not just a chip capacity problem but also a debt financing problem. The institution estimates that by 2029, AI-related outstanding debt could exceed $70 trillion, and the cumulative AI capital expenditure from 2024 to 2029 could reach around $11.1 trillion. The annual AI capex includes GPUs, networking, storage, supporting CPUs, and data centers to accommodate this compute, possibly surpassing $2 trillion by 2028.

Within this framework, for any AI compute project to materialize, three things need to come together: capital, offtake, and data center. Neocloud Company needs a loan to purchase GPUs, but banks usually require to see long-term offtake contracts first; customers want to see GPU and data center capacity locked in before committing; data center operators are more willing to lease capacity to stronger credit hyperscalers such as Microsoft, Google, Amazon.

NVIDIA is getting involved in this cycle. SemiAnalysis states that NVIDIA is starting to provide a minimum revenue guarantee, i.e., backstop, for some of Neocloud's GPU leasing revenue. For lending institutions, this embeds NVIDIA's credit into the project financing; for Neocloud, it reduces the difficulty of financing; for NVIDIA, it is not just selling GPUs but also sharing in the leasing revenue above the guaranteed minimum price. The institution estimates that under this structure, NVIDIA's average take rate may be around 18%-20%.

This makes NVIDIA increasingly resemble a credit hub in the AI era. By guaranteeing and directly leasing data center capacity, it helps Neocloud address capital, offtake, and data center constraints simultaneously. If this structure expands, the buyer base in the AI compute market will no longer be limited to a few hyperscalers and large AI labs; more enterprises, inference service providers, and small and medium-sized AI-native companies may also gain access to compute power.

JPMorgan Chase

The hashrate financing addresses whether demand can be unleashed, while TSMC addresses whether the most advanced chips can be manufactured. In its report on July 7, JPMorgan Chase maintained an Overweight rating on TSMC and raised its target price from NT$2,500 to NT$3,100. The bank raised TSMC's EPS forecasts for 2026, 2027, and 2028 by 5%, 10%, and 16% respectively, citing stronger AI demand, tighter advanced process capacity, higher gross margins, and increased visibility on price hikes in 2027.

JPMorgan Chase expects TSMC's gross margin in the second quarter of 2026 to approach 70%, reaching 69.5%; even with the dilution from N2 ramp-up and overseas fab expansion, the gross margin in the second half of 2026 and the first half of 2027 could still be maintained in the high 60% range. The bank also raised TSMC's data center AI revenue compound annual growth rate for 2024-2029 from 59% to 69%, and anticipates that demand for AI CPUs, ASICs, network chips, and accelerators will collectively drive TSMC's revenue.

In terms of capacity, JPMorgan Chase projects TSMC's capital expenditures to reach $58 billion, $78 billion, and $84 billion in 2026, 2027, and 2028 respectively, totaling about $219 billion over three years. N3 capacity is expected to reach 167,000, 213,000, and 240,000 wafers per month by the end of 2026, 2027, and 2028 respectively; N2 capacity could reach 170,000 wafers per month by the end of 2028, faster than previous nodes. Advanced packaging is also accelerating, with CoWoS annual capacity possibly approaching 2 million wafers in 2027, benefiting from demand for AI accelerators, Vera, Venice, and other products.

Morgan Stanley

If TSMC represents the manufacturing bedrock of AI computing, Samsung and SK Hynix represent the other end of the memory cycle. Morgan Stanley commented on Samsung Electronics' preliminary second-quarter results on July 7, stating that the company's revenue is approximately ₩171 trillion, a year-on-year increase of 129%; operating profit is around ₩89.4 trillion, a year-on-year increase of 1,810%, broadly in line with market expectations. Memory profit could be close to ₩92 trillion, primarily driven by the strong prices of DRAM and NAND; the Foundry/LSI losses have narrowed to about ₩2 trillion.

Morgan Stanley maintained an Overweight and Top Pick rating on Samsung, with a target price of ₩381,000, implying about a 20% upside from the closing price on July 6. The bank believes Samsung is still in a steep profit recovery cycle, and although the stock has soared this year, it has not yet fully reflected the extent of the memory profit rebound. Long-term agreements and the advantage in advanced DRAM nodes may also enhance earnings predictability.

However, Morgan Stanley also warned in another semiconductor industry brief that the short-term outlook for memory stocks may be cooling off. The bank believes that memory is still in a structurally bullish market driven by AI capital spending, with earnings expected to grow by 35%-40% by 2027; but indicators such as year-over-year price increases, inventory improvements, and breadth of earnings upgrades are approaching their peak. In other words, the cycle is not over, but the rate of change is reaching its peak, and crowded positions need to be digested.

This is also the core contradiction of current AI trading: the fundamentals are still strong, but popular assets have accumulated significant long positions. Morgan Stanley stated that the market is debating whether large cloud providers have excess AI computing power. If hyperscalers maintain or increase capital spending, a pullback in memory stocks may provide a buying opportunity; if they cut spending, the narrative of excess computing power will continue to suppress the sector. The bank is more inclined towards DRAM and traditional memory in the short term, believing they are superior to NAND, and is least optimistic about memory module manufacturers.

Nomura Securities

Outside of computing and memory, optical interconnect has become another highly resilient segment. Nomura Securities reiterated a Buy rating on InnoLight on July 6 and raised the target price from RMB 1,015 to RMB 1,325, implying a potential upside of about 20.6% from the closing price on July 6. Nomura believes that the recent pullback in AI infrastructure stocks has not altered InnoLight's growth drivers for 2026-2028.

The firm is optimistic about the expansion of the 1.6T optical module, silicon photonics, NPO, and CPO markets, and has raised its global high-end optical module shipment forecast. Nomura expects that the shipments of 800G optical modules in 2027/2028 will reach 55 million/78 million units, 1.6T shipments will reach 71.5 million/126 million units, and will begin to contribute to 2.4T and 3.2T products. InnoLight's share in the global AI data center optical module market is expected to remain at 30%-35%.

Based on this, Nomura has raised its revenue forecast for InnoLight in 2027-2028 by 28%-37% and profit forecast by 30%-38%. The firm believes that high-end optical communication products are still one of the bottlenecks in AI data centers, with supply chain shortages expected to persist in the short term, but from 2028 onwards, as suppliers in China, Japan, and the US expand production, some bottlenecks may ease.

The demand for AI is even filtering down to less visible material segments. In another report on Japanese chemicals and textiles, Nomura stated that AI servers are driving demand growth for high-end MLCCs, and the trend towards multilayered MLCCs will increase the usage of release film and base film. The firm expects that material demand will grow at an average annual rate of about 10% from the benchmark year of 2025 to 2028.

MLCC is a key passive component in servers, communication equipment, and electronic terminals. As AI servers require higher-capacity and higher-stability electronic components, the internal electrode layers of MLCCs increase, leading to a higher demand for high-precision release film materials during the production process. Nomura believes that Japanese companies dominate this niche market, with Lintec, Toyobo, Toray Industries, and Mitsui Chemicals all poised to benefit.

Specifically, Nomura rates Lintec as a Buy due to its leading market share in the MLCC release film market and its strong position in high-end applications such as AI servers. While Toyobo has a Neutral rating, its Utsunomiya plant's new production line is expected to contribute significantly by the third quarter of 2026, with a potential increase in market share post-2027. Toray focuses more on the upstream base film, with Nomura estimating its market share in related export base film markets to exceed 50%.

These reports together illustrate the second phase of the AI investment theme. The first phase was identifying who could directly sell the scarcest GPUs, HBM, and servers; the second phase requires understanding how capital enters the system, how capacity expands, how orders are secured, and how demand spills over into optical modules, advanced packaging, passive components, and materials.

This also signifies a shift in investors' selection criteria. Mere labeling with AI is no longer sufficient, as the market is beginning to question where each company lies in the chain: is it a bottleneck asset or a replaceable supplier? Does it have pricing power or is it at the mercy of upstream price hikes? Is its growth driven by real CapEx or by short-term inventory and reorders? Has its valuation already reflected the most optimistic earnings revision?

The conclusion is not a cooling of AI CapEx but rather a more discerning market. Nvidia's backstop shows that the computational power demand is strong enough to reshape the credit market; TSMC's capacity expansion and price hikes indicate that advanced processes are still in short supply; Samsung and memory stocks' profit recovery demonstrate that AI is still driving the hardware cycle; Unimicron and the MLCC material chain indicate that demand is spreading deeper into the supply chain.

However, the same set of reports also serves as a reminder that the hotter the sector, the more likely temporary crowding becomes. The next phase of AI trades may no longer simply chase the largest market leaders but instead seek out specific positions within financing, manufacturing, memory, optical interconnect, and materials that still face capacity constraints, earnings upgrades, and valuation mismatches.