The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Ten semiconductor companies drew attention in 2025 for bets on AI infrastructure beyond the GPU: optical links, edge inference, cluster networking, analytics processing and hardware security. The list below reconstructs CRN’s mid-2025 selection, not a ranking of the strongest investments or a claim about which companies later succeeded. “Hottest” here means a mix of funding, product announcements and strategic activity reported at the time—not proof of revenue, broad availability or customer adoption.
That distinction matters to investors and technology buyers alike. A funding round can extend a company’s runway; a product launch can show technical progress. Neither alone establishes production shipments, repeat customers or product-market fit. The ten businesses also sell different things, so their milestones are best compared within their categories.
CRN’s list and its reported 2025 milestones are the basis for this retrospective. Read CRN’s original 2025 list.
Why AI infrastructure created room for more than GPU startups
Large AI systems depend on more than processors. Data must move between compute and memory, and among machines in a cluster; those transfers can affect power use, latency and the number of accelerators a system can keep busy. Other workloads have different constraints: edge devices need efficient inference in limited space and power, analytics systems may benefit from specialized processing, and distributed infrastructure needs security that begins below the operating system.
#1 Best Overall
- COMPATIBILITY: Development board supporting multiple wireless protocols including Bluetooth
- Thread, Matter, Zigbee, ANT, and NFC at 2.4GHz frequency
- PROCESSOR: Features the advanced nRF54L15 transceiver chip from Nordic Semiconductor for reliable wireless communications
- WIRELESS STANDARDS: Implements IEEE 802.15.4 protocol support for Matter, Thread, and Zigbee networking applications
- DEVELOPMENT PLATFORM: Comprehensive evaluation board designed for testing and prototyping wireless connectivity solutions
The companies on CRN’s list address different pieces of that problem. Some build accelerators, others develop optical connectivity or networking systems, and one focuses on silicon root-of-trust technology. They are not ten interchangeable alternatives to Nvidia. Their prospective value depends on where they fit, whether software and system integration are workable, and whether customers can deploy them at scale.
How to read the 2025 list
CRN did not publish a transparent scoring method for “hottest,” so the selection should be treated as editorial rather than as a formal ranking. Its reported signals include funding rounds, product launches, roadmaps and partnerships. These are useful indicators of attention and company activity, but they answer different questions.
- Funding indicates that investors committed capital, not that a company has revenue or profitable demand.
- A launch or announcement does not by itself establish that a product is shipping broadly, production-qualified or orderable.
- A partner network can widen routes to market, but does not establish deployment volume.
- Performance figures may refer to a particular workload, configuration or company comparison. Peak bandwidth and TOPS are not the same as end-to-end application performance.
Availability and commercial progress described below are limited to the milestones CRN reported in 2025. The cited coverage does not establish what happened after that snapshot.
Rank #2
- 【ACEBOTT ESP32 Development Board】 - Powerful WiFi and wireless development board, driven by the rugged ESP 32 module, seamlessly integrated with Arduino IDE. With Hall sensors, high-speed SDIO/SPI, UART, I2S and I2C, it is the cornerstone of IoT and smart home innovation.
- 【Wi-Fi/Bluetooth and Arduino Cloud Compatibility】 - This board uses 2.4GHz dual-mode WiFi and wireless chips with low-power technology, which are RoHS-compliant, simplifying wireless communication and allowing you to easily connect devices and platforms. Whether you are using a compatible Arduino IDE or exploring other development environments, our board can easily adapt to your needs.
- 【Improved and Professional Edition】 - All IO pins are brought out for easy development; no additional breadboard is required; the Type-C interface is equipped with electrostatic discharge protection diodes and transient voltage suppression diodes to protect the chip from damage by electrostatic breakdown and various surge pulses. In addition, it is equipped with a freeRTOS operating system, which is very suitable for the Internet of Things, smart homes, and building smart robots/game consoles.
- 【Easy to Use】- The ACEBOTT ESP-32 Development Board includes everything you need to support the microcontroller. Just connect it to a computer via a USB cable or use an AC-DC adapter or battery to power it to start using it. Whether you are an experienced developer or a hobbyist, this development board can provide you with the tools you need for unlimited innovation.
- 【 Install Plugins And Download Drivers】: This ESP32 development board includes detailed instructions on how to download plugins and all necessary programs and codes from the network environment. The path is: ACEBOTT official website - Resources - WIKI.
The 10 companies, grouped by what they build
| Company | Primary area | Reported 2025 milestone | Key adoption question |
|---|---|---|---|
| Ayar Labs | Optical I/O and chiplets | $155 million funding round; UCIe-compatible optical chiplet announced, with up to 8 Tbps claimed | Can optical I/O be integrated into customer packaging and manufacturing workflows? |
| Axelera AI | Edge AI acceleration | Partner Accelerator Network launched; EU support of up to €61.6 million reported for Titania | Do partner channels lead to deployments, and does its software support target workloads? |
| Celestial AI | Photonic connectivity | $250 million Series C1; more than $515 million in total funding reported | Can its Photonic Fabric clear manufacturing, qualification and system-integration hurdles? |
| Cornelis Networks | AI and HPC networking | CN5000 400-Gbps family launched; faster products described as roadmap items | Will customers adopt another networking stack alongside established options? |
| EnCharge AI | Low-power AI inference | $100 million Series B; EN100 accelerator launched | How do its precision-specific figures translate to customers’ models and software? |
| Lightmatter | Silicon photonics and optical interconnect | Passage M1000 and L200 announced; CRN reported $850 million backing and a $4.4 billion valuation | Can package-level optical bandwidth deliver useful system-level gains at scale? |
| Speedata | Analytics acceleration | Callisto APU and C200 PCIe card launched; $44 million Series B | Do customers’ database operations map to the accelerator’s supported workload? |
| Tenstorrent | AI processors, software and IP | More than $693 million Series D reported; Blackhole PCIe cards launched | Can its software and ecosystem attract users beyond early adopters? |
| Xsight Labs | DPUs and Ethernet switching | E1-SoC and E1-Server announced, the latter described as an 800G DPU system | What production-ready software and system support accompany the hardware? |
| zeroRISC | Hardware security | $10 million oversubscribed seed round; OpenTitan-based security focus | How will OEMs validate and integrate the commercial security offering? |
Optical connectivity: Ayar Labs, Celestial AI and Lightmatter
Ayar Labs is pursuing optical I/O: moving data optically between chips or components to address bandwidth and power constraints associated with electrical connections. CRN reported an announced UCIe-compatible optical interconnect chiplet with up to 8 Tbps of claimed bandwidth and a $155 million funding round. Compatibility with UCIe may help align the technology with a chiplet-based design approach, but it does not remove the need to solve packaging, thermal, manufacturing and system qualification challenges. The cited coverage does not establish broad shipment or customer deployment. Ayar’s overview describes its positioning: Ayar Labs’ company overview.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Celestial AI describes its Photonic Fabric as a platform for connectivity, switching and packaging, aimed at linking processors and memory in AI systems. CRN reported a $250 million Series C1 and total funding above $515 million. That capital signals investor backing for a technically demanding effort; it does not settle how quickly the technology can be manufactured and qualified. Bandwidth, latency, energy and cost advantages need to be assessed against specific architectures and workloads rather than inferred from a platform description.
Lightmatter is also targeting optical connectivity and package-level interconnect. CRN reported its Passage M1000 and L200 announcements, $850 million in backing and a $4.4 billion valuation in 2025. It reported company claims of 114 Tbps total optical bandwidth for M1000, more than 200 Tbps total I/O bandwidth per package for L200, and up to eight-times-faster training in a specified comparison. Those are not interchangeable measures: aggregate bandwidth does not establish application throughput, and a training-speed comparison depends on the model, system configuration and baseline. The valuation is a reported 2025 figure, not a current valuation.
All three companies work around optical data movement, but their technologies and intended system roles are not identical. Buyers evaluating them would need to determine whether the proposed component fits a chiplet, package or broader system design, and what integration and supply-chain commitments it requires.
AI acceleration: Axelera AI, EnCharge AI and Tenstorrent
Axelera AI is focused on edge inference rather than directly competing for hyperscale model training. CRN reported a Partner Accelerator Network with more than 15 partners, including Lenovo, Dell Technologies, Advantech and Arduino, as well as EU support of up to €61.6 million for its Titania chiplet. “Up to” is a maximum, not evidence that the full amount was paid; the cited article does not specify disbursement or conditions. Its Metis platform and partner relationships suggest a route into embedded and industrial systems, but partner participation alone does not show deployment scale. Buyers should check whether its software stack supports the exact models and frameworks they use.
EnCharge AI uses analog in-memory computing for inference and launched the EN100 accelerator, according to CRN, alongside a $100 million Series B. The coverage reported an M.2 version with up to 200 TOPS in an 8.25-watt envelope and a PCIe version with four NPUs and approximately one quadrillion operations per second. These are product figures, not a direct measure of application speed: precision, model, workload and power-measurement conditions affect comparisons. A low-power design may appeal for local or edge inference, but software porting and workload fit remain central evaluation questions.
Rank #4
- The C8051F320 /1 series utilizes the proprietary CIP-51 microcontroller core of Silicon Labs. The CIP-51 is fully compatible with MCS-51M instruction sets; Software can be developed using standard 803x / 805x assembler and compiler
- The CIP-51 core provides all the peripherals that come with the standard 8052, including four 16-bit counters/timers, full-duplex UART with extended baud rate configuration, enhanced SPI ports, 2304-byte on-chip RAM, 128-byte Special Function Register (SFR) address space and 25/21 I/0 pins.
- 10-Bit ADC, Up to 200 ksps, Up to 17 or 13 external single-ended or differential inputs ,VREF from external pin, internal reference, or VDD
- USB specification 2.0 compliant, Full speed (12 Mbps) or low speed (1.5 Mbps) operation, Voltage Regulator Input: 4.0 to 5.25 V
- C8051F320 Single Chip Development Board built-in temperature sensor, External conversion start input, Two Comparators, Internal Voltage Reference, POR/Brown-Out Detector
Tenstorrent combines AI processors with software and IP licensing, rather than relying on a single hardware-sales model. CRN reported more than $693 million raised in a Series D, a $2 billion pre-money valuation, Blackhole PCIe card launches and a partnership with AIREV for an enterprise and sovereign generative-AI stack. Its interest in open software and licensing offers a different proposition from a tightly integrated incumbent ecosystem, but “open” should not be read to mean every hardware design, firmware component and software tool is equally open under the same terms. The cited coverage does not establish software maturity or broad adoption; buyers should validate compiler, framework, model and support coverage before committing workloads.
Networking and DPUs: Cornelis Networks and Xsight Labs
Cornelis Networks sells into AI and high-performance computing networking. CRN reported the launch of its 400-Gbps CN5000 family, based on Omni-Path, and a roadmap for an 800-Gbps CN6000 in 2026 and a 1.6-Tbps CN7000 in 2027. Roadmap targets are not shipping products. Cornelis also claimed, in specified comparisons, two-times-higher message rates, 35% lower latency and up to 30% faster HPC performance than Nvidia’s 400-Gbps InfiniBand NDR, plus six-times-faster collective communication than RoCE for AI applications. These company-reported comparisons should not be treated as universal results without the test conditions, workload and system configuration. The practical question is whether performance or cost advantages outweigh the expense of adopting and operating another fabric. Cornelis’ newsroom provides its company announcements: Cornelis Networks newsroom.
Xsight Labs announced the Arm-based E1-SoC and E1-Server, described by CRN as an 800G DPU system. A DPU can offload or combine infrastructure functions such as networking, storage and security; programmability is useful only if customers can deploy, monitor and support the necessary software. Xsight’s claims about control-plane and data-path programmability do not alone establish production readiness or parity with mature alternatives such as Nvidia BlueField or AMD Pensando. The cited coverage does not detail customer deployments or the maturity of its software ecosystem.
Best Value
- [Open Source Resources] Coming with extensive open-source documentation and example projects.
- [Integrated Design] Seamlessly integrates with MaixPy for a powerful and compact system.
- [Supportive Community] Active forum and telegram channel for user support and collaboration.
- [Gowin 51 | R 20k | FPGA Technology] Utilizes cutting-edge FPGA technology, enabling rapid prototyping and customization.
- [Extended Temperature Range] Operates stably at 105℃, perfect for high-temperature applications.
Analytics acceleration: Speedata
Speedata targets database and big-data operations with its Callisto Analytics Processing Unit and C200 PCIe card, launched according to CRN. The company also raised $44 million in a Series B. CRN reported a 280-fold performance improvement for a pharmaceutical workload versus a non-specialized processing unit. That should be read as a reported result for a particular workload, not a general speedup across databases or customer environments. The card’s value depends on how well real queries map to the operations it accelerates and whether adoption requires changes to existing pipelines. The cited article does not establish broad availability or the representativeness of the benchmark.
Hardware security: zeroRISC
zeroRISC focuses on a silicon root of trust and device-management software, built around the OpenTitan project. CRN reported a $10 million oversubscribed seed round. The underlying problem is important as infrastructure becomes more distributed: systems need a way to establish trust and manage devices below the operating system. The commercial question is what zeroRISC adds through integration, silicon IP, management or services, and how OEMs validate it against their threat models. The cited coverage does not establish broad adoption or specify the licensing and commercial scope of every component.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What buyers and investors should verify
Comparing these companies by a single headline number is misleading. A buyer considering an accelerator, network fabric or optical component should test it against the intended system and workload, not against unrelated peak figures. For an investment or vendor review, separate the evidence into technical capability, ability to deploy, and evidence of demand.
- Product stage: Is the offering an architecture announcement, sampling part, evaluation card, shipping product or volume deployment? A launch announcement does not answer all of these.
- Workload evidence: For performance claims, identify the model or query, precision, baseline, system configuration and whether the measurement is chip-level or end-to-end.
- Software fit: Check compilers, drivers, framework integrations, model portability, observability and production support. Silicon specifications cannot compensate for missing tooling.
- Integration burden: Optical chiplets and package-level systems may require advanced packaging and foundry coordination; a new accelerator or network may require software and operational changes.
- Commercial proof: Distinguish investors and ecosystem partners from disclosed customers, recurring deployments, shipment volumes and revenue. The CRN coverage does not supply enough evidence to infer those outcomes for every company.
- Execution exposure: Foundry access, advanced packaging capacity, long qualification cycles, capital needs and incumbent bundles can all affect the path from a promising design to a sustainable business.
The larger question behind the list
The 2025 group illustrates two possible routes to semiconductor opportunity. One is to replace an incumbent processor or network in a defined workload. The other is to improve the interconnect, packaging, networking, analytics or security layers that let existing processors work more effectively. Which route produces durable companies depends less on a striking funding total or peak specification than on whether a product solves a costly customer problem and can be integrated, supported and supplied reliably.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
