HomeArticle

In the AI era, FPGA is on the rise again.

半导体行业观察2026-11-16 09:06
The rapid expansion of AI data centers is bringing new market opportunities for FPGA.

The rapid expansion of AI data centers is bringing new market opportunities for FPGAs.

On November 2, Lattice released its financial results for the third quarter of fiscal 2026, with revenue reaching 291 million U.S. dollars, a year-on-year increase of 118%; among which FPGA revenue increased by about 90% year-on-year. Ford Tamer, CEO of the company, said that the acceleration of AI demand in data centers is an important factor driving business growth.

This set of data has brought FPGAs back into the industry's field of vision. In the past, FPGAs were once highly expected to expand heterogeneous computing, but in the rapid development of the AI computing market, they have never become the mainstream computing platform.

Nowadays, the growth of AI data centers is opening up another path. GPUs are undertaking more and more core computing tasks, but the continuous expansion of the scale of servers, cabinets and clusters has also made system management, interface control and security protection more complex. The programmable feature of FPGAs can just play a role in these links.

From competing for computing cores to entering the underlying layer of the system, FPGAs have ushered in a new development period in the development of the AI industry.

Giants placed big bets, but FPGAs failed to enter the mainstream

Semiconductor giants once hoped that FPGAs could become an important part of future computing architectures.

In 2015, Intel announced the acquisition of Altera for approximately 16.7 billion U.S. dollars. At that time, Intel hoped to combine FPGAs with Xeon processors to provide data centers with more flexible heterogeneous computing solutions. Through CPU-FPGA collaboration, users can perform hardware acceleration for specific workloads to achieve higher performance and energy efficiency beyond general-purpose processors.

Coincidentally, AMD also took a fancy to this market. In 2020, AMD announced the acquisition of Xilinx for approximately 35 billion U.S. dollars, and completed the transaction in 2022. Compared with the CPU-FPGA collaboration that Intel emphasized at that time, AMD hoped to use Xilinx to complement its adaptive computing capabilities and incorporate CPUs, GPUs, FPGAs and related software technologies into a more complete product system.

At that time, the mainstream judgment of the industry was that data processing demands were becoming more and more diversified, and a single architecture could not cover all scenarios. FPGAs, which can flexibly adjust hardware logic according to application requirements, were expected to become an important part of heterogeneous computing. Therefore, behind the two acquisition cases lies the common expectation of giants for the future computing architecture.

However, the development of the AI industry has gradually changed the competition pattern of the computing market.

Large model training and inference have brought huge parallel computing demands. With its mature software ecosystem, developer base and large-scale deployment capabilities, GPUs have quickly become the core computing platform of AI data centers. At the same time, ASICs optimized for specific workloads are also expanding their applications. The former occupies the mainstream AI computing market with the advantages of versatility and ecosystem, while the latter pursues higher performance and energy efficiency in specific scenarios through customized design.

Sandwiched between the two routes, FPGAs are in an awkward situation. In the face of general-purpose computing, FPGAs cannot compete with the mature software ecosystem and large-scale capabilities of GPUs; in the face of special-purpose computing, ASICs can carry out targeted optimization around fixed workloads. Although FPGAs have the advantages of programmability, low latency and flexible adaptation, the high development threshold, difficulties in toolchain and IP ecosystem construction also limit their large-scale applications.

This also prevented Intel and AMD from fully realizing their original vision for FPGAs. Subsequent business adjustments also made the follow-up of the two acquisitions seem somewhat anticlimactic.

In 2025, Intel announced the sale of 51% of Altera's equity to Silver Lake, and Altera was transformed into independent operation. Data shows that Altera achieved revenue of 1.54 billion U.S. dollars in fiscal 2024, but its operating loss under GAAP reached 615 million U.S. dollars. Although AMD retained Xilinx's product and technology system, FPGAs have never become the core of its AI computing business.

Therefore, FPGAs need to find a new market foothold, and the continuous expansion of AI data centers just provides such a soil.

AI data centers bring new positions for FPGAs

The new space opened up by AI data centers for FPGAs first comes from the changes of the computing system itself.

With the continuous growth of AI computing power demand, data centers are moving from construction centered on computing chips to complex systems with collaboration of multiple links such as computing, network, power supply and heat dissipation. GPUs are responsible for processing core computing tasks, but to make large-scale computing power run stably, a large number of hardware modules need to cooperate with each other. As a result, the connection and management between servers, between boards, and between different functional modules have become more complex.

This change has gradually highlighted the demand for hardware control beyond computing. Taking AI servers as an example, as the deployment scale expands from a single machine to a full cabinet and cluster, the number of GPUs increases, and high-speed networks, power supply and heat dissipation systems are becoming increasingly complex. Servers not only need to monitor temperature, voltage and equipment status in real time, but also coordinate communication between different modules, and handle tasks such as power-on, reset, fault response and secure boot.

Here we have to mention AMD again. Although the original intention of acquiring Xilinx has not been realized, it has found a more suitable application position for FPGAs. AMD applies FPGAs to server motherboard management. Its Spartan UltraScale+ series is oriented to data center server I/O and motherboard management controller (BMC), and is responsible for power supply sequence management, temperature and voltage monitoring, fan control, I/O expansion and system status monitoring on the server motherboard. In June 2026, AMD announced that the Spartan UltraScale+ SU200P had entered the mass production preparation stage, and was officially mass-produced in July of the same year.

Compared with CPUs and GPUs, these tasks do not require complex general-purpose computing capabilities, but require hardware to respond in a timely manner and run continuously. FPGAs can undertake related work as a dedicated board-level controller. As the number of sensors, power domains and interfaces integrated on the server motherboard increases, the role of FPGAs in underlying management and control is also increasing.

On the other hand, Lattice chooses to extend the application of FPGAs from server motherboards to security management and cabinet control. At the server level, Lattice's low-power FPGAs can undertake tasks such as secure boot, hardware identity authentication, firmware protection, sensor monitoring and interface management; at the cabinet level, they can cooperate with management software to monitor server operating status, power consumption and temperature, helping data center operators carry out centralized management. In October 2026, Lattice and Arm announced a cooperation to combine Lattice FPGAs, Arm AGI CPU server platforms and AMI firmware.

Different from the previous two companies, the independent Altera further extends FPGAs to the data path of data centers. In DPUs and AI network cards, FPGAs can undertake tasks such as RDMA, virtual switching, network security and data preprocessing; in storage systems, they can be used for NVMe-oF protocol processing, data encryption and compression. If these tasks are handed over to the CPU for processing, certain computing resources will be occupied. By offloading some functions to FPGAs, the CPU load can be reduced, and programmable logic can be used to adapt to different network protocols and system architectures.

It can be seen that the opportunities that AI data centers bring to FPGAs are not limited to increasing computing capabilities. With the expansion of computing power scale, data centers have higher requirements for underlying hardware control, equipment management and system collaboration. For FPGA manufacturers, these previously relatively marginal links are becoming new market entrances.

FPGA manufacturers start to compete for platform entrances

The entry of FPGAs into AI data centers has not only brought changes in the number of application scenarios, but also changed the competition mode between manufacturers.

In the past, FPGA manufacturers mainly competed around chip performance, logic resources, power consumption and development tools. But in the AI data center market, whether products can enter server design also depends on their collaboration capabilities with processors, firmware and management software. Platform support and ecosystem construction beyond chips are becoming new focuses for manufacturers to compete for the market.

In July 2026, Lattice completed the acquisition of AMI. AMI has long provided BIOS, BMC firmware and data center infrastructure management software, with its business covering server boot, hardware management and platform operation and maintenance. The most critical point is that AMI can provide software and firmware support connecting processors and underlying hardware. For Lattice, this transaction complements its originally relatively weak software and firmware capabilities, and also gives its FPGA business the opportunity to enter the more underlying management architecture of servers.

Different from Lattice's acquisition to complement software and firmware capabilities, Altera focuses more on covering diverse system needs by enriching the FPGA product portfolio. In October 2026, Altera released the Agilex 2X product portfolio, targeting fields such as digital AI infrastructure, networking and physical AI. Its product layout covers connection capabilities such as high-speed transceivers, Ethernet and PCIe, and also involves applications such as board-level and cabinet control, interface bridging and hardware root of trust.

At the same time, AMD is also continuing to rely on Xilinx's technical accumulation to promote the layout of adaptive computing and embedded products. It can be seen that although the competition paths of various manufacturers are different, the common point is that FPGAs are extending from a single chip product to a wider range of system applications.

In addition to product and business layout, FPGA manufacturers have also begun to participate in the construction of data center management standards.

In October 2026, Lattice and AMI promoted the establishment of the SCMP working group under the Open Compute Project (OCP), with the participation of AMD, Arm, Intel, Jabil, ASPEED, Microsoft and other enterprises. The working group hopes to promote an open, vendor-neutral security control and management architecture to improve the compatibility of identity authentication and management interfaces between different devices.

As the scale of data centers expands, the management relationship inside servers and between different devices becomes more and more complex. A unified management standard helps reduce the difficulty of integration between hardware from different manufacturers, and also reduces the dependence of platform construction on specific vendor solutions. For chip manufacturers participating in it, standard formulation is also related to the right to speak of future platform architecture and the opportunity for their own products to enter the industrial chain.

Can China's FPGA industry seize new opportunities

Overseas FPGA manufacturers are accelerating the expansion of their businesses to AI data centers, and Chinese manufacturers are also facing a new market opportunity.

In terms of market size, Grand View Research predicts that China's FPGA market will grow from about 2.101 billion U.S. dollars in 2023 to 4.689 billion U.S. dollars in 2030, with a compound annual growth rate of 12.2% from 2024 to 2030. Communications and industrial control are still important application fields, while data centers, automotive electronics and artificial intelligence provide new growth directions for the market.

However, the expansion of market demand does not mean that domestic FPGAs can directly share the dividends of AI computing power growth. Especially in the data center field, FPGA manufacturers need to improve product capabilities for different application scenarios. High-performance computing and high-speed network applications require chips to have strong logic resources, high-speed SerDes, advanced interfaces and power control capabilities; server management pays more attention to low power consumption, interface compatibility, security control and long-term stable operation capabilities. Different application directions put forward differentiated requirements for chip performance, function configuration and supporting ecology.

In the previous text, the layout of overseas manufacturers provides a referable idea: domestic FPGAs do not need to concentrate all resources in the high-end computing acceleration market, and can also find breakthroughs from more specific system requirements. For example, links such as server board-level management, interface expansion, protocol conversion and security control have actual demands for the flexible adaptation capability of FPGAs. Compared with directly participating in the high-performance computing competition around GPUs, these segmented scenarios may be more suitable as the starting point for some domestic manufacturers to enter the data center supply chain.

However, entering the data center supply chain requires not only hardware products, but also development tools, IP resources, software support and platform adaptation capabilities. If domestic manufacturers want to further move from providing chips to platform introduction, they also need to make up for these links.

At present, the advantage of the domestic market lies in its huge scale of server and data center market, rich application scenarios, and local FPGA manufacturers also have certain flexibility in customer communication, customized development and supply chain collaboration. If product development can be carried out around the actual needs of local server manufacturers, and cooperation can be established with processors, BMC and firmware suppliers, there is an opportunity to shorten the product verification and introduction cycle.

Of course, these advantages will eventually be tested by products and the market.

Conclusion

The renewed attention of FPGAs in the AI era reflects that the demand of data centers for hardware capabilities is becoming more diversified. The future competition takes place not only in the computing core represented by GPUs, but also in the underlying links that support the stable operation of the entire system.

For FPGA manufacturers, the value brought by this round of changes lies in expanding the application space of products, and making software and hardware collaboration, platform adaptation and ecological construction become new competition links. FPGAs may not become the protagonist of AI computing, but they have the opportunity to occupy a more important position in AI infrastructure.

This article is from the WeChat official account "Semiconductor Industry Observation" (ID: icbank), author: Xia Xue, published with authorization from 36Kr.