ASIC Mega-Deals Land in Quick Succession: Broadcom Arranges Over $50 Billion for OpenAI as Custom Chip Supply Chain Roles Shift

The custom chip (ASIC) industry has seen several financing deals worth tens of billions of dollars in less than a month. These deals are also changing how chipmakers and their customers work together.

According to a report by 36Kr, Broadcom is arranging more than $50 billion in financing for OpenAI's custom AI chips. Talks are still at an early stage. Less than a week earlier, a financing deal worth up to $42 billion between Anthropic and Broadcom was disclosed. This money is intended to pay for the rental of computing power using TPUs. Two leading AI model companies are now linking multibillion-dollar financing deals to custom chips. One is moving toward using its own chips, while the other is expanding its use of rented computing power built around custom chips.

There is also a lot of activity in chip design. Synopsys signed a multiyear agreement worth more than $1 billion with Amazon for chip design technology. SEMIFIVE won a $52 million contract to develop an AI inference ASIC. On October 6, Marvell raised its revenue target for its custom chip business in fiscal 2029 to more than $12 billion. Spending on custom computing is spreading beyond the main AI accelerator chips to include design services, reusable chip designs and supporting chips.

Financing deals bring chip suppliers closer to their customers' finances

The Anthropic-Broadcom deal is linked to computing costs that Anthropic has already agreed to pay. Documents disclosed on October 1 show that Anthropic has committed to about $125.2 billion in TPU computing leases over five years. The financing line of up to $42 billion could cover about one-third of that amount. Broadcom or its chosen financing partners will provide the money as convertible debt, which can be changed into shares under certain conditions. Anthropic is not expected to issue the related notes until after it completes an initial public offering (IPO). The credit line does not mean Broadcom has already spent $42 billion in cash. But it means the relationship between the two companies now goes beyond simply buying and supplying chips.

The TPUs involved in this deal were developed jointly by Google and Broadcom. In April this year, Anthropic announced that it would begin obtaining several additional gigawatts of TPU computing power from 2027. The financing deal provides a way to pay for that expansion. For Google, TPUs are moving beyond use in its own services and are also serving the large computing needs of outside AI model companies. This creates room for more shipments. Anthropic, meanwhile, can increase its use of custom chips without first having to finish developing its own chip.

Independent AI accelerator makers are also finding new customers. On September 29, Cerebras and General Compute announced a multiyear agreement. General Compute will buy, install and operate Cerebras systems, then sell dedicated AI inference computing services to customers. These services are expected to begin in the first quarter of 2027. The first target is coding agents. These programs repeatedly use AI models to plan, write code, test it and make changes. Even a short wait for each model response can add up over the course of a whole task.

General Compute had previously secured a $400 million debt financing line from Upper90. This is not the value of the chip contract. However, it shows that new types of cloud service providers are taking on the costs and work of financing and running hardware. This allows companies that provide AI inference services to buy computing power without having to own all the equipment themselves. Independent chipmakers gain another way to reach customers, without requiring every application company to buy hardware directly.

Just one day earlier, Cerebras and Gimlet Labs announced plans to deploy 100 megawatts of AI inference computing power. The first partner data center is expected to begin operating later this year. Gimlet runs a platform for AI inference that supports several types of chips. The two companies had already worked together to handle customers' private AI inference tasks. These two partnerships, announced close together, bring specialized accelerators into cloud services. They also show that the market can grow through channels beyond the custom chip systems built for internal use by companies such as Google and Amazon.

Forecasts based on chip shipments are not the same as actual revenue

An industry research report released in late September predicts that ASIC shipments will exceed GPU shipments in the high-end cloud AI accelerator market in 2027. It forecasts that about 25.95 million high-end cloud AI accelerators will be shipped that year. Of these, about 15.26 million will be ASICs, an increase of about 111% from the previous year and nearly 60% of the total. However, these figures count chips, not their revenue or computing power. They do not prove that ASIC revenue or computing power will exceed that of GPUs. The same forecast expects shipments related to Nvidia to grow by about 18%. If demand grows enough, both types of chips can expand at the same time.

Segment	Recent major deal/partnership	Amount or scale	Timing
Financing arrangement	Broadcom arranges financing for OpenAI custom chips	Over $50 billion	Disclosed in early October
Financing arrangement	Broadcom-Anthropic convertible debt financing	Up to $42 billion	Disclosed October 1
IP licensing	Synopsys-Amazon multi-year IP agreement	Over $1 billion	September 30
Design services	SEMIFIVE takes on inference ASIC development for US chip company	About $52 million	September 29
Guidance	Marvell raises fiscal 2029 custom chip revenue target	Over $12 billion	October 6

Note: Most of these amounts are maximum contract values or targets. They do not equal confirmed revenue or money already paid.

Large TPU deployments still depend on a small number of customers. New partnerships for independent accelerators also need time to prepare data centers and move customers onto the services. Financing deals and long-term agreements bring expected demand into the chip supply chain earlier. But continued purchases will depend on whether the services using this computing power can attract enough customers to use it.

In-house chips move from design plans to server racks

OpenAI is moving its own chip project closer to actual use. In a late-September interview, hardware chief Richard Ho said that Jalapeño will first meet OpenAI's own computing needs and that selling it to outside customers is not currently a priority. Details released on October 2 show that this custom accelerator uses AMD's EPYC Turin as its host CPU, the main processor that supports the accelerator.

Under the announced division of work, Broadcom handles chip implementation and networking-related tasks, while Celestica helps put the boards and server racks together. The AI model company can design accelerators around its own needs while continuing to use established CPUs and specialist suppliers to build the systems. Expanding its own chip development does not mean it has to build every part of a server from scratch.

When testing Jalapeño, OpenAI compared how much energy it used to run AI inference while meeting the same response-speed requirements. Reducing the movement of data was an important design goal. Because OpenAI controls its models, software and real-world usage requirements, it can adjust the chip design to match changes in computing work and memory access during inference. Ho also stressed that Jalapeño can be programmed to run different models. Earlier tests included models with openly available weights, such as DeepSeek and Kimi. Jalapeño is not permanently designed to run just one model. However, supporting new model families still requires engineers to develop computing instructions called kernels and keep improving performance.

On September 25, Amazon announced a specific example of adapting a model to its chips. The real-time interactive AI platform Reactor worked with the AWS Neuron team to run the video generation model Rolling Forcing on a single Trainium2. They achieved real-time generation at more than 16 frames per second. Unlike systems that create a whole video at once, this type of application must keep producing frames and respond to user input. It must keep delays low throughout the generation process. The team developed special computing kernels to handle changes in data size, memory access and cache management. It also changed how the model's work was divided among the chip's cores. The results did not reveal the scale of commercial deployment, and they cannot be directly used to prove that the system costs less than one based on GPUs. However, they show the work needed to bring ASICs into new applications. Chipmakers must work with model and platform developers to fix problems that arise during actual use before a chip's expected performance can become a useful service.

The Gimlet-Cerebras partnership takes a different approach by combining different types of chips. Gimlet's platform can assign tasks to different chips at each stage of AI inference. This lets GPUs and Cerebras systems handle the work each is best suited for, while the platform provides the service through one interface. Application developers can add specialized accelerators behind the scenes without moving their entire service away from their existing GPU systems at once. Dedicated chips may first take over part of the inference process rather than replace an entire group of servers.

Under plans OpenAI announced earlier, Jalapeño is expected to begin deployment by the end of the year. The company has also made clear that it will continue to use accelerators from Nvidia and other partners. Recent projects involving in-house chips, adapting models for cloud chips and combining different types of chips all focus on matching hardware to particular tasks. GPU suppliers now face customers who have more ways to decide how to spend their next round of chip purchases. ASIC suppliers, meanwhile, must prove they can deliver reliable results and win ongoing business, not just show strong performance in a single test.

Chip design service providers are joining projects earlier

On September 29, SEMIFIVE announced a contract worth about $52 million to develop an AI inference ASIC for an unnamed US chip company that designs chips but does not manufacture them. This is SEMIFIVE's largest single contract so far and its first North American project under a model in which the customer gives the requirements and the supplier takes responsibility for the later development work. The customer sets the product requirements and performance targets. SEMIFIVE handles detailed chip design, software, packaging, testing and mass production. In the traditional model, the customer completes most of the design before asking a service provider to help turn it into a working chip. Under the new model, the supplier joins earlier and takes on more responsibility. Companies without large engineering teams can keep control over what their products should do and how well they should perform, while leaving the technical implementation to specialists.

The inference chip will use LPDDR6 memory and a PCIe interface, which connects it to other computer components. Its design goals include lowering memory power use and the total cost of owning and operating the system. The design is scheduled to be sent for manufacturing in the first half of 2027. Mass production is planned for 2028. There is still a long way to go from signing the contract to shipping chips. SEMIFIVE has won both an order and responsibility for the later testing, manufacturing and delivery work.

On September 30, Synopsys announced a multiyear agreement worth more than $1 billion with Amazon. The deal focuses on chip IP designed for specific uses. IP means reusable chip design blocks. They let cloud providers avoid developing every part from scratch, so they can spend more of their own engineering resources on chip designs and software that make their products stand out. Synopsys specifically said the partnership would move its IP business toward a model that combines license fees with royalties based on the number of chips sold. Instead of earning fees only during the design stage, suppliers may continue to earn money as customers increase production. The agreement did not reveal which specific products it covers. Amazon's custom chips also serve general computing, AI and networking needs, so the full $1 billion cannot be counted as orders for AI inference chips.

Marvell's higher forecast on October 6 shows another source of growth. The company raised its fiscal 2029 custom chip revenue target to more than $12 billion. Management said the increase mainly came from supporting chips used alongside accelerators. These opportunities involve several customers and projects, rather than one main chip for one customer. Marvell's custom chip opportunities include network interfaces, security, memory expansion, storage and chips that move inference tasks away from the main processor. As accelerator systems grow, standard components may not meet customers' needs for bandwidth, network setups and system management. Cloud providers are therefore asking for more custom designs for these supporting parts. Marvell uses technology from its existing networking, storage and other products to create partly customized or fully custom designs based on customer needs. This means the main accelerator is not the only part of the system creating new chip design projects.

This also explains why companies that provide chip IP and design services can win orders even as customers invest more in their own chips. Customers mainly want control over product requirements, chip architecture and software. They may not want to build full engineering teams to develop every connection or system management chip. However, revenue from supporting chips and shipment numbers for high-end AI accelerators measure different things. Marvell's higher revenue target does not directly prove that demand for main ASIC chips is growing by the same amount.

For now, ASIC growth is bringing in more types of buyers and changing how work is shared among suppliers. AI model companies are taking more control over product design, new types of cloud service providers are handling deployment, and specialist suppliers are taking on more design and mass-production work. After the big deals are signed, only solutions that can be deployed on time, keep adapting to new models and win orders for future generations will turn this investment into a lasting share of the market.