Anthropic's trillion-dollar listing stirs up the semiconductor world
Semiconductor industry vertical and horizontal
2026.11.22 14:16
· From Beijing
Full text 4711 words
Text | Semiconductor industry vertical and horizontal
On November 19, Anthropic, a large artificial intelligence model company, was officially listed on Nasdaq with the stock code "ANT". The IPO was finally priced at US$83 per share, higher than the previous recommendation range of US$70 to US$78. It issued 980 million shares and raised approximately US$81.34 billion, exceeding the global IPO fundraising record set by SpaceX in June this year. The opening price on the first day of listing rose to US$111.50 and closed up by about 17%. The company's valuation reached the trillion-dollar level.
Different from conventional technology stock IPOs, this listing simultaneously launched a US$26 billion private placement, with almost all subscribers coming from Anthropic's upstream computing power and semiconductor supply chain: Google and Nvidia each invested US$10 billion, AMD invested US$2.5 billion, Broadcom, Micron and Microsoft each invested US$1 billion. TSMC, which has always maintained strict neutrality and rarely directly invests in downstream software companies, also invested US$500 million through related entities. All seven companies subscribed for non-voting Class N ordinary shares at the issue price.
Upstream hardware suppliers collectively invest to become new shareholders of downstream customers without requiring voting rights, which was rare in past semiconductor industry cycles. Combined with the US$900.9 billion in computing power and infrastructure commitments disclosed in Anthropic’s prospectus, this listing is not only a capital event, but also directly exposes changes in AI chip competition, upstream supply chain procurement methods, and computing power financing models.
From single GPU to multi-chip parallelism
The direct reason why semiconductor companies are intensively participating in the placement is that the top large model buyers have spread their computing power purchases across multiple hardware routes. Anthropic disclosed in the prospectus submitted to the U.S. Securities and Exchange Commission that the training and inference of its Claude series models run simultaneously on Nvidia GPU, AMD GPU, Google TPU and Amazon AWS Trainium, and is working with Broadcom to promote the deployment of customized AI accelerator XPU and supporting network equipment.
For chip manufacturers other than Nvidia, Anthropic is a key customer to verify large-scale commercial deployment. In the field of general-purpose GPUs, public information shows that Anthropic is deploying up to 2 GW of AMD MI450 and Helios systems. Coupled with OpenAI's $90 billion GPU commitment, AMD has obtained structural orders in the data center market that do not rely on Nvidia's capacity overflow. In this placement, AMD invested US$2.5 billion, second only to NVIDIA among chip design companies, and is intended to stabilize long-term cooperation with this core customer.
Cloud service providers' self-developed chips also rely on Anthropic to achieve scale. Amazon AWS's Project Rainier cluster has deployed more than 1 million Trainium2 chips for it, becoming the world's only chip platform besides Nvidia and Google to achieve ultra-large-scale cutting-edge model training. In April this year, Anthropic promised AWS to spend at least US$110 billion between May 2026 and April 2036, corresponding to up to 5 gigawatts of computing power. Google’s cumulative commitment reached US$121.3 billion, and the two parties expanded their TPU cooperation with Broadcom to 3.5 GW in April.
Broadcom, the leader in customized chips, has benefited significantly from this round of multi-chip competition, but it also faces structural concerns. In the third quarter of fiscal year 2026, Broadcom's AI semiconductor revenue increased 221% year-on-year to approximately US$16.7 billion, and fourth-quarter guidance was raised to US$21.7 billion. Management expects annual AI chip revenue to exceed US$100 billion in fiscal 2027, and plans to deliver more than 20 gigawatts of XPU and Ethernet solutions by 2028. However, the risk of concentration of a single customer always exists. In June this year, Macquarie downgraded Broadcom's rating due to concerns that MediaTek would cut into Google's TPU design. It is expected that Broadcom's share of Google's customized chips may drop from about 95% in 2026 to 65% in 2028.
Directly accepting Anthropic's customized XPU orders and shipping Tomahawk 6 switching chips and 200G SerDes high-speed interconnect products has become a key increment for Broadcom to hedge against Google's share diversion. However, many Wall Street analysts also pointed out that in order to meet the needs of large customers, Broadcom is getting more involved in the delivery of complete machine racks. Although reselling server hardware can expand the scale of revenue, it will directly lower the overall gross profit margin.
The increased volume of customized ASICs has not replaced the position of general-purpose GPUs in cutting-edge research and development, but it has effectively weakened the bargaining power of a single architecture. Industry estimates show that ASICs optimized for specific models have a total cost of ownership that is 30% to 50% lower than general-purpose GPUs. They have obvious performance advantages per watt and are suitable for large-scale reasoning with relatively stable structures. However, ASICs are tightly coupled with existing algorithms, model iterations can easily reduce hardware efficiency, and they lack a software ecosystem similar to CUDA. Therefore, Anthropic still signed an NVIDIA GPU cluster agreement of up to US$120.5 billion through SpaceXAI, with the term extended to 2029 and a fully loaded monthly average expenditure of approximately US$3.1 billion. However, the agreement allows both parties to terminate the contract 90 days in advance after the initial non-cancellable period in January and March 2027. This shows that top buyers are not willing to lock in excessive long-term exposure on a single GPU architecture; Nvidia invested US$10 billion to subscribe for non-voting shares. Although it cannot influence customer selection, it can maintain software compilation and deep collaboration on the next-generation platform in a multi-chip parallel structure.
Direct mining model is transmitted to the upstream
Bigger supply chain changes are occurring among chip designers. In the past, wafer foundries, advanced packaging and storage manufacturers mainly connected with chip design companies and cloud manufacturers, but Anthropic's "Anthropic Direct" direct procurement strategy disclosed in the prospectus is shortening this link. According to a report quoted by TrendForce on November 18, Anthropic is bypassing the traditional packaged procurement model of public cloud and directly leasing AI accelerators, CPUs, HBM and network equipment to external parties, while also separately contracting data center computer rooms and operation and maintenance.
The logic of this split comes from the cycle mismatch between the two types of infrastructure. Data center civil engineering, power access and liquid cooling facilities often take several years to build, while the iteration cycle of AI chips and HBM has been compressed to 12 to 18 months. By signing separate contracts for space and computing hardware, buyers can lock in scarce power in advance, defer expensive chip and storage purchases until a new generation of products becomes available, and at the same time compress the intermediate markups from cloud vendors.
Anthropic is not the only one doing this. The same report from TrendForce pointed out that Google is also evaluating the direct purchase of HBM for its own TPU to reduce the cost markup caused by passing through custom chip design partners such as Broadcom. TrendForce predicts that the total capital expenditures of the world's nine largest cloud service providers will increase by 106% year-on-year to US$957 billion in 2026, and will increase by another 60% to exceed US$1.5 trillion in 2027. In the face of trillions of dollars in capital expenditures, a few percentage points of intermediate cost compression corresponds to an expenditure difference of tens of billions of dollars. It is difficult to reverse the trend of large buyers directly connecting with upstream original manufacturers.
TSMC broke with convention and invested in downstream model companies, which is happening at this node. On November 10, TSMC's board of directors approved a one-time capital budget of approximately US$40.93 billion for advanced processes, advanced packaging and factory construction. It also approved the establishment of a semiconductor ecological investment company of up to US$500 million, and invested up to US$500 million in new Anthropic shares and OpenAI preferred shares respectively. On November 19, TSMC disclosed that it had completed the subscription of approximately 6.024 million Anthropic shares. The two equity investments totaling US$1 billion accounted for only 2.44% of the current capital budget and were obviously not motivated by financial arbitrage.
Combined with TSMC's current production capacity investment, the intention of this investment is very clear. In the third quarter of this year, TSMC's revenue reached NT$1.494 trillion, equivalent to approximately US$46.72 billion, of which 3 nanometers accounted for 33% of wafer revenue, and 2 nanometers accounted for 8%. In October, consolidated revenue hit a new high of NT$586.89 billion, a year-on-year increase of 59.7%, and annual capital expenditures were raised to US$60 billion to US$64 billion. As a key bottleneck connecting computing chips and HBM, CoWoS advanced packaging has been directly written into the risk warning of Anthropic's prospectus; TSMC plans to increase CoWoS prices by 10% to 20% and expand monthly production capacity from over 130,000 pieces in 2026 to approximately 200,000 pieces in 2027. Both 2nm and CoWoS are heavy asset investments with payback periods of several years. The demand forecasts delivered by chip designers often have an amplification effect. The core of TSMC's investment in "customers of customers" is to use very small funds in exchange for channels that can penetrate the middle layer and directly check the true rhythm of terminal computing power.
Micron Technology invested US$1 billion to participate in the placement, and its logic is consistent with TSMC. With the popularity of long-context reasoning and large-parameter clusters, the capacity and bandwidth of HBM have become the core indicators restricting system efficiency. Micron CEO Sanjay Mehrotra has previously advanced his forecast for the HBM market to reach US$100 billion from 2028 to 2027, and bluntly stated that even if new fabs increase volume from 2028, it is still difficult to judge whether supply can catch up with demand. The company recently acquired a 300mm wafer factory in Taiwan, China, with a clean room of approximately 300,000 square feet. It is expected to achieve effective shipments in fiscal year 2028 and simultaneously promote the second phase of construction. After the direct procurement model brings HBM into the direct planning vision of large model companies, storage manufacturers are no longer just secondary suppliers passively waiting for GPU manufacturers to place orders, but have begun to intervene in the definition of hardware specifications for AI systems earlier.
Capital structure changes and long-term risks
This round of AI semiconductor expansion has been accompanied by a core controversy: How much of the rapid growth of chip manufacturers is supported by their own advance capital or credit guarantees? Less than a month before Anthropic went public, on October 26, Broadcom, together with Bank of America, Blackstone, Citigroup and Morgan Stanley, announced a US$63 billion financing transaction. A new special purpose entity will purchase Broadcom XPU and network equipment and lease them to Anthropic starting in 2027. Three banks arranged first-lien financing, with Blackstone exclusively providing second-lien funding and Broadcom providing residual value support for the first-lien portion.
Broadcom's regulatory filings the next day disclosed specific risk exposures. After all racks are deployed, Broadcom's maximum potential liability on an undiscounted basis will be approximately US$43 billion. Although this is not a cash payment in the current period, after the hardware is delivered, the supplier still needs to bear joint and several liability for the customer's performance and changes in the residual value of the equipment for a long period of time. In addition, Amazon and Google have previously invested more than US$30 billion in total, and the total stock market value corresponding to the issue price of US$83 exceeds US$307.9 billion. This time, the seven major suppliers have added an additional US$26 billion in placements. The market's doubts about "circular financing" continue to heat up before the listing.
Anthropic completed a record listing, which did improve buyers' payment base in the short term. Including the US$131.85 billion in cash and short-term investments on the books as of September 30, the company's pro forma cash reserves after the listing and placement are expected to exceed US$220 billion. Gil Luria, director of technology research at D.A. Davidson, pointed out in an interview on November 19 that this cash scale clearly differentiates Anthropic from new cloud service providers such as CoreWeave. The financial report shows that CoreWeave's revenue in the third quarter of this year was US$3.661 billion, but its capital expenditure reached US$14.511 billion. Its operating cash flow could only cover 21%. Its expansion is still highly dependent on Nvidia and external bond market support.
Joey Brookhart, head of the Tokenomics team at SemiAnalysis, estimates that Anthropic will need approximately US$280 billion in cumulative revenue in the next four years to cover its computing power bill. Based on the company's revenue growth rate of US$33.24 billion in the first nine months of this year and US$17 billion in the third quarter, the annualized rate in a single quarter has reached US$68 billion. As of the end of October, the run rate was approximately US$81 billion. This threshold is covered. After listing, even though Anthropic has not yet obtained an investment-grade credit rating, it has opened up the public debt market and stock payment channels, and the necessity for chip manufacturers to use their own credit to guarantee tens of billions of dollars in residual value, like Broadcom, will be significantly reduced.
But for the semiconductor supply chain, risks have only been transferred, not eliminated. Not all of the US$900.9 billion computing power commitment disclosed in the prospectus is locked: approximately US$451.6 billion is due before the end of 2030, and approximately US$70.7 billion is due in the next 12 months. Another US$186.2 billion comes with enforceable termination rights or preconditions. Once the counterparty's financing, computer room construction or approval is blocked, the relevant contract can be withdrawn.
The pressure on the cost side is also clearly visible. In the third quarter, Anthropic’s computing power and infrastructure expenditures rose to US$9.9 billion, accounting for 58% of revenue from 48% in the second quarter, of which US$5 billion was short-term computing power expenditures whose prices were significantly higher than long-term contracts, which directly caused GAAP operating profit to shift from a profit of US$974 million in the second quarter to a loss of US$567 million. In addition, about 79% of the company's revenue comes from consumption-based businesses billed based on Token usage, and most customers have not signed long-term lock-in contracts. Gil Luria pointed out that even if Anthropic loses in the model competition in the future, as long as the overall AI usage continues to grow, cloud giants and chip factories can still take on the computing power needs of alternative models; however, for wafer factories, packaging and testing factories, and storage factories that have invested tens of billions of dollars in expanding production, once there is a mismatch between the pace of upstream capacity release and the downstream business realization cycle, performance pressure will still be quickly transmitted to the upstream.
Conclusion
Seven supply chain giants spent US$26 billion to subscribe for non-voting shares, putting the interests of the AI semiconductor industry chain on the table. Anthropic's listing will provide additional cash for computing power purchases in the next few years, and will further accelerate multi-chip competition and upstream direct procurement. However, the open market only solves the problem of whether payment can be made at the moment. How much of the US$900.9 billion contract can ultimately be realized, and whether the new production capacity of 2 nanometers and HBM can be smoothly digested, depends on the actual returns of downstream applications.
Respect for the original work, if you are qualified, you should be appreciated.
According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.
Express
13:45
Six companies have launched A-share IPO coaching, covering physical AI, commercial aerospace, and semiconductor materials.
13:43
The Xujiapo Tunnel of the Kangyu Section of the west-Chongqing High-speed Railway was successfully completed
13:41
Liu Liehong, Director of the National Data Administration, went to Jiangxi Province to investigate the progress of the reform of market-oriented allocation of data elements
13:39
The war affects energy supply, Iran implements power restrictions
13:12
The second domestically produced large-scale cruise ship "Aida Huacheng" officially launched its maiden commercial voyage
12:29
Chen Wenqing will go to Germany to attend the 6th China-Germany High-Level Security Dialogue and visit Germany and Austria
12:16
New AI method may help optimize sepsis emergency treatment process
12:15
China's first batch of innovative development pilot projects for trusted data space have made phased progress
12:00
The AI cheating incident in South Korea's college entrance examination is spreading, and the Minister of Education resigns.
11:58
XPeng Huitian’s first full-size, full-tilt electric flying car successfully completed its tilting transition test flight
11:56
Trains suspended in Xigang section of Qinghai-Tibet Railway have resumed operation, and the line section has been comprehensively inspected
Four new species of large fungi discovered in Junzifeng, Fujian
11:28
JiKr’s Legal Department refutes rumors about “recall” text messages: it’s an impersonation scam
11:27
Lang Ping won the FIVB's highest honor, the "Volleyball Medal"
11:08
People with the same name and surname lead to mispronunciation, Sohu Zhang Chaoyang refutes the rumor that "a warning letter was issued"
11:02
Xie Jun, deputy general manager of China National Petroleum Corporation, under review and investigation
11:00
Multiple space science experiments (experiments) on the China Space Station are being advanced in an orderly manner
10:54
Amazon buys Israeli AI chip startup for nearly $7.5 billion, aiming to take on Nvidia
10:36
Since November, more than 60 A-share companies have received special loan commitment letters for stock buybacks and shareholding increases.
10:19
Guo Xiangyu, general manager of China Chengtong, went to Chengdu and Chongqing to conduct work research
Scan to download App

