AI networking: which startup is ahead?

Why is AI networking suddenly such a big market?

AI networking is growing incredibly fast because connecting GPUs has become one of the biggest challenges in building larger AI systems.

The latest figures from IDC show just how quickly spending is increasing. Worldwide data center Ethernet switching revenue reached $12.3 billion in the second quarter of 2026, up 64.5% from a year earlier. Nvidia generated $2.5 billion in quarterly Ethernet switch revenue, growing 181.1%, while Arista reached approximately $2.5 billion in total switching revenue, up 37.6%.

The shift toward faster connections is equally striking. According to IDC, 800GbE equipment represented 35.8% of data center Ethernet switching revenue in the first quarter, rising to 41.2% in the second. That is a gain of more than five percentage points in just three months. Combined with 200G and 400G equipment, high-speed connections now account for roughly three-quarters of data center switch spending.

The reason is straightforward. Thousands of GPUs working on the same AI model constantly exchange information. Slow connections leave expensive processors waiting, wasting computing capacity. Faster and better-managed networks help companies get more useful work from the GPUs they already own.

This has created an opening for startups such as DriveNets, Nexthop AI, Ayar Labs and Lightmatter. They are trying to solve different parts of the same problem, and their progress varies considerably.

What does it mean to lead the AI networking market today?

The leading AI networking startup should have technology that customers are actually deploying, a growing business and a credible advantage over existing suppliers.

There are two main groups competing for that position. DriveNets and Nexthop AI build networking systems that connect GPUs and servers across AI data centers. Ayar Labs and Lightmatter develop optical technology designed to move information between processors and networking equipment faster and more efficiently.

These technologies overlap in purpose, but they solve different problems. A company building optical connections directly into chip packages cannot be compared fairly with one selling complete Ethernet networking systems using a single bandwidth figure. Optical chips and Ethernet systems are different bets, so we cover which technologies are gaining traction in our report covering AI infrastructure.

We also need to distinguish three types of connections. Scale-up networking connects accelerators working closely together. Scale-out networking connects servers and racks across a large computing cluster. Scale-across networking links infrastructure in separate data centers.

Commercial adoption remains our most important test. A company operating networks for paying customers has stronger evidence of leadership than one announcing impressive laboratory performance or a multibillion-dollar valuation.

Optical networking deserves to be examined separately because the technology is still moving toward widespread production. In that market, manufacturing readiness, working integrated systems and semiconductor partnerships tell us more than current revenue alone.

Which AI networking startups are leading right now?

DriveNets is currently ahead in commercial AI networking, while Ayar Labs leads optical interconnect development and Nexthop AI is the strongest emerging challenger in complete networking systems.

DriveNets has disclosed more than $1 billion in secured business, positive cash flow since 2025 and commercial AI networking deployments. Nexthop raised $500 million in March 2026 at a $4.2 billion valuation, but has published less comparable financial information.

The optical competition is closer. Ayar Labs raised $650 million in primary financing during 2026 and demonstrated an integrated optical computing system with Wiwynn and Global Unichip. Lightmatter has a reported $4.4 billion valuation and demonstrated 1.6 terabits per second per fiber with technology developed alongside Qualcomm.

Those achievements tell different stories. DriveNets has the strongest disclosed commercial record. Nexthop has attracted extraordinary backing for a young company. Ayar Labs is progressing toward optical production, while Lightmatter is pushing the limits of photonic connectivity.

Here is how we rank the main contenders based on the evidence currently available.

Startup	Main technology	Strongest evidence	Our assessment
DriveNets	Ethernet AI fabrics	$1B+ secured business; commercial deployments	Overall commercial leader
Nexthop AI	AI Ethernet systems	$500M Series B; $4.2B valuation	Strongest emerging systems challenger
Ayar Labs	Optical I/O	$650M raised in 2026; integrated system demonstrations	Optical commercialization leader
Lightmatter	Photonic interconnects	$4.4B valuation; 1.6 Tbps per fiber	Advanced optical contender
Xsight Labs	Switching and DPU silicon	$300M+ financing; operator design selections	Credible chip challenger
Upscale AI	AI networking silicon and systems	~$500M raised; Nvidia investment	Well-funded emerging competitor
Avicena	microLED optical links	1 Tbps evaluation kits	Earlier-stage optical specialist

Is DriveNets actually selling AI networking technology at scale?

DriveNets has the strongest disclosed commercial business among the independent AI networking startups we examined, although its $1 billion in secured business includes more than its newer AI products.

In June 2026, DriveNets announced a $410 million Series D, bringing cumulative funding to $1 billion. More interestingly, the company disclosed over $1 billion in secured business and said it had been cash-flow positive since 2025. Rounds of this size raise the question of who is getting funded, and by whom, which we look at in our report on AI infrastructure.

That commercial position took years to build. Founded in 2015, DriveNets initially focused on large telecommunications networks. Its technology was adopted by operators including AT&T and Comcast, giving the company experience operating infrastructure where reliability and traffic management are critical.

It has since expanded into AI infrastructure. In May 2025, WhiteFiber announced a deployment of DriveNets AI Fabric for GPU-as-a-Service infrastructure. A year later, DriveNets made its Fabric Scheduled Ethernet solution available through Dell AI Factory, opening another route to customers building AI systems.

There is one important catch: DriveNets has not disclosed how much of its secured business comes specifically from AI networking. Its established telecommunications operations contribute to the company's financial record.

Still, positive cash flow, existing network operators and named AI customers put DriveNets in a stronger commercial position than younger competitors whose public disclosures focus mainly on financing and product launches.

Has DriveNets proved its AI networking works beyond one data center?

DriveNets has already demonstrated a commercially deployed AI network connecting GPU clusters 52 miles apart, one of the strongest practical achievements among networking startups today.

In July 2026, DriveNets and WhiteFiber announced Project Redwood, connecting two data centers equipped with Nvidia H200 GPUs into one logical supercluster.

The companies reported 111.2 terabits per second of validated bandwidth and 0.9 milliseconds of latency across approximately 83 kilometers of dark fiber. WEKA provided the storage technology involved in the project.

The distance is particularly interesting. Large AI facilities consume enormous amounts of electricity and require substantial cooling capacity. Connecting facilities in different locations allows infrastructure operators to expand without concentrating every GPU inside the same building.

There are limits to what Project Redwood proves. The disclosed latency and bandwidth figures describe a particular network configuration. They do not establish that every AI training workload will perform equally well across geographically separated facilities.

What makes the deployment compelling is its commercial progression. WhiteFiber first announced DriveNets AI Fabric for its infrastructure in 2025, then expanded the relationship into a multi-site supercluster in 2026. We have a named customer moving toward a more demanding network architecture, with measured performance rather than a laboratory demonstration alone.

Could Nexthop AI overtake DriveNets?

Nexthop AI is probably DriveNets' most dangerous young competitor, although its $4.2 billion valuation is currently backed by less publicly disclosed commercial evidence.

Founded by former Arista Networks executive Anshul Sadana, Nexthop raised $110 million before announcing a $500 million Series B in March 2026. The round was led by Lightspeed Venture Partners, with Andreessen Horowitz joining as a major investor.

That brought reported cumulative funding to $610 million. The Series B alone accounted for approximately 82% of the company's total disclosed financing, showing how quickly investors increased their commitment.

Nexthop has also introduced switches designed for scale-out, scale-across and front-end networking in cloud and AI data centers. Its products address the same kinds of infrastructure requirements that have made Nvidia's networking business so successful.

The company benefits from Sadana's experience at Arista, one of the world's leading data center networking suppliers. But Nexthop still needs to prove that it can turn a strong team, considerable funding and new equipment into large recurring customer orders.

DriveNets disclosed more than $1 billion in secured business alongside its latest financing. Nexthop has not published a directly comparable figure for revenue, customer commitments or profitability.

For now, DriveNets has the better commercial record. Nexthop could change that ranking quickly if its new networking systems gain traction with major cloud providers.

Is Ayar Labs ahead of everyone in optical AI networking?

Ayar Labs currently has the strongest case for leadership in optical AI scale-up connections, thanks to its manufacturing progress, semiconductor partnerships and increasingly complete system demonstrations.

Its funding history shows how quickly the company has advanced. Ayar reported $370 million in cumulative financing after its December 2024 Series D. It raised another $500 million in March 2026, followed by an additional $150 million in September.

Those two rounds brought its primary financing during 2026 to $650 million and its disclosed cumulative funding to approximately $1.02 billion, excluding separately disclosed strategic investments where amounts were unavailable. Ayar raised $650 million in one year; we sort out who is ahead and who is falling behind in our report on who's ahead in AI infrastructure.

Ayar's technology centers on TeraPHY optical engines and SuperNova light sources. Together, these components aim to move information through optical connections close to computing chips, reducing some of the limits associated with longer electrical connections.

The company's progress is increasingly visible at the system level. In October 2026, Wiwynn, Ayar Labs and Global Unichip announced a live integrated optical computing demonstration. It combined GUC semiconductor technology with Ayar's optical components and included real-time monitoring of link health and temperature.

Compared with demonstrations showing only that an optical connection can carry data, this is a meaningful step toward equipment that customers can operate inside AI infrastructure.

Ayar has also attracted strategic investment from companies including AMD, Intel, Nvidia, MediaTek and Wiwynn. That gives it relationships across several parts of the semiconductor and server ecosystem.

We still lack public production revenue and shipment figures on the scale needed to establish mass adoption. But among independent optical interconnect companies, Ayar has assembled the strongest combination of financing, hardware integration and manufacturing preparation.

Is Lightmatter's optical technology better than Ayar Labs'?

Lightmatter has demonstrated exceptional optical bandwidth density, but Ayar Labs currently looks slightly closer to large-scale commercial deployment.

In March 2026, Lightmatter announced that its Passage co-packaged optics chiplet had successfully demonstrated 1.6 terabits per second per fiber. The technology uses 16 optical wavelengths and incorporates Qualcomm's high-speed optical SerDes technology.

Lightmatter described the result as an eightfold improvement in bandwidth density compared with selected existing near-packaged and co-packaged optics solutions.

Its development goes beyond one performance record. The company introduced Passage L20 technology designed to reduce fiber requirements and developed a detachable optical connector intended to simplify manufacturing and maintenance. It has also worked with GUC, Cadence and Synopsys on semiconductor design and integration.

In June 2026, Lightmatter announced that it had joined Nvidia's NVLink Fusion ecosystem, positioning its technology for future AI computing platforms. Ayar Labs also has close relationships with Nvidia and other accelerator manufacturers.

Comparing the two companies through raw bandwidth would be misleading. Lightmatter's 1.6 Tbps figure measures the capacity of a fiber connection under a specific architecture. It says little on its own about the total performance, power consumption or economics of a deployed AI computing system.

Ayar's recent work with Wiwynn and GUC provides somewhat stronger evidence of system-level readiness. Lightmatter may have the more ambitious optical architecture, but we would currently give Ayar the commercialization advantage.

Which AI networking startups have raised the most money?

Ayar Labs and DriveNets are the best-funded independent AI networking companies in our comparison, with approximately $1 billion each in announced cumulative financing.

Lightmatter follows with approximately $850 million, while Nexthop AI has raised $610 million. Upscale AI reached about $500 million less than two years after its founding.

Across six companies, Ayar Labs, DriveNets, Lightmatter, Nexthop AI, Upscale AI and Xsight Labs, we identified at least $4.3 billion in announced financing. The first three account for approximately two-thirds of that total.

The pace is almost as striking as the amounts. DriveNets raised its $1 billion over roughly a decade. Nexthop assembled $610 million much faster, while Upscale AI reached $500 million within approximately 18 months.

Investors are funding several competing approaches to the same infrastructure problem, from Ethernet systems to silicon photonics. Building these products is expensive: advanced chip development, manufacturing qualification, equipment inventories and customer support require substantial investment before large revenues appear.

Valuations tell a different story. DriveNets was reportedly valued at approximately $8.5 billion in June 2026, almost twice Nexthop's valuation and more than twice Ayar's March valuation. DriveNets has more disclosed commercial evidence to support that premium, although we cannot determine whether the price is reasonable without annual revenue figures.

For perspective, an $8.5 billion valuation would represent 17 times annual revenue at $500 million, or 42.5 times annual revenue at $200 million. These are hypothetical examples, since DriveNets has not published the revenue figure needed to calculate its actual multiple.

Startup	Announced cumulative funding	Latest reported valuation	Main market
Ayar Labs	~$1.02B	$3.75B	Optical I/O
DriveNets	$1.0B	~$8.5B	Ethernet fabrics
Lightmatter	~$850M	$4.4B	Photonic interconnects
Nexthop AI	$610M	$4.2B	AI networking systems
Upscale AI	~$500M	$2.0B	AI networking silicon and systems
Xsight Labs	$300M+ in latest round	$2.8B	Switch and DPU silicon
Avicena	~$120M	Not publicly established	microLED optical links

The funding figures reflect announced financing, rather than current cash balances. For Xsight Labs, the figure shown represents its latest disclosed round rather than a verified lifetime total.

Can Xsight Labs and Upscale AI challenge the bigger networking startups?

Xsight Labs and Upscale AI are becoming serious competitors in AI networking hardware, although neither has demonstrated the commercial reach of DriveNets.

Xsight Labs announced a financing round exceeding $300 million in July 2026, reaching a $2.8 billion post-money valuation. The company said multiple global network operators had selected its E1 800G data processing unit and X2 switch. A $2.8 billion valuation means more with context, so we lay out what these companies are worth in our AI infrastructure report.

Xsight develops chips that other companies can incorporate into networking equipment. Its technology could reach numerous customers through original equipment manufacturers without Xsight having to sell complete data center networks itself.

Upscale AI is pursuing a broader approach, combining networking silicon, systems and software. The company emerged from stealth in September 2025 with $100 million in funding. By June 2026, it had raised approximately $500 million and reached a $2 billion valuation.

Its investors include Nvidia, Temasek and Salesforce Ventures. Upscale's Skyhammer architecture targets fast communication between AI accelerators, memory and storage.

The two companies represent different opportunities. Xsight already has announced product selections from network operators. Upscale has gathered substantial financial backing for an ambitious AI-specific networking architecture.

Neither has released enough information on production shipments or recurring revenue to demonstrate market leadership. Nevertheless, Xsight's commercial design selections and Upscale's unusually rapid financing make both worth watching closely.

Will optical networking replace Ethernet in AI data centers?

Optical interconnects will probably replace more copper connections inside AI systems, but Ethernet will remain central to how data centers communicate.

The distinction comes down to what each technology does. Ethernet controls how information moves through a network. Optical interconnects use light to transmit that information through physical connections.

Modern Ethernet networks already use optical fiber extensively. Companies such as Ayar Labs and Lightmatter want to bring optical connections much closer to processors and switches, where electrical connections increasingly struggle with distance, bandwidth and power consumption.

That creates opportunities for several technologies to coexist. Nvidia and Broadcom can continue selling networking equipment while incorporating new optical components. Ayar Labs can provide optical I/O engines, and Lightmatter can supply photonic interconnect technology.

Avicena is pursuing another route through microLED-based optical connections. In August 2026, it announced the shipment of 1 Tbps evaluation kits, providing customers with hardware to test its approach.

The likely result is an AI networking market where Ethernet systems, optical components and advanced semiconductor packaging develop together. Optical startups can become major suppliers without replacing the networking protocols that their customers already use.

Are AI networking startups taking market share from Nvidia?

AI networking startups are winning customers, but Nvidia is still expanding so quickly that we have no convincing evidence of a significant market-share shift toward independent challengers.

IDC's latest figures show Nvidia generating approximately $2.5 billion in data center Ethernet switching revenue in the second quarter of 2026, up 181.1% year over year. Nvidia held 20.4% of the segment, remaining its largest branded supplier.

Arista represented another 18.7% of data center Ethernet revenue, while Cisco held 18.2%. Together, these three established companies accounted for approximately 57% of the market. Three established names hold about 57% of Ethernet switching; we look at how far the challengers are behind in our report on AI infrastructure.

The scale is worth considering. Nvidia generated roughly $4.6 billion in Ethernet switching revenue across the first two quarters of 2026, using IDC's reported quarterly figures. That is more than four times DriveNets' entire announced cumulative fundraising.

The comparison mixes revenue with financing, so it cannot measure profitability or relative business value. It does show the enormous operating scale of the companies startups are trying to challenge.

Nvidia also maintained its leadership while the market grew. Its data center Ethernet revenue increased from $2.1 billion in the first quarter to $2.5 billion in the second, a gain of approximately 19% quarter over quarter.

We can see opportunities for independent companies in multi-vendor networks, specialized optical connections and custom switching silicon. But Nvidia's recent sales give us little reason to believe that its position is already weakening.

Does DriveNets have an advantage over Nvidia in open Ethernet networking?

DriveNets has a credible advantage for customers who want to combine AI accelerators from different suppliers, especially those trying to avoid depending entirely on Nvidia's networking stack.

Nvidia sells tightly integrated systems combining GPUs, Spectrum-X Ethernet switches, BlueField data processing units and networking software. That integration offers customers a well-supported architecture with components designed to operate together.

DriveNets takes a more flexible approach. Its AI Fabric supports Ethernet-based networks built from equipment supplied by multiple vendors, making it attractive to operators using heterogeneous accelerator clusters.

AMD's involvement is particularly relevant. AMD joined DriveNets' June 2026 financing, and the companies subsequently published an AI networking reference architecture built around AMD Instinct MI350-series accelerators.

In July, DriveNets reported approximately 5% higher throughput and 10–15% lower time to first token in selected comparisons using its technology. These figures come from specific reference configurations and publicly available benchmark comparisons. They should not be interpreted as proof that DriveNets universally outperforms Nvidia in identical tests.

The commercial argument is clear. Large AI operators may want to combine AMD GPUs, different switches and specialized networking software as they expand their infrastructure.

DriveNets has positioned itself around that demand, with AMD as both an investor and technology partner. Its challenge is to demonstrate that this flexibility consistently produces better economics for customers than tightly integrated alternatives.

Are Ayar Labs and Lightmatter close to mass production?

Ayar Labs and Lightmatter are much closer to manufacturing-ready optical AI networking than they were two years ago, but neither has publicly established widespread production deployment across major AI data centers.

Ayar Labs has made the clearest transition toward complete systems. Its work with Wiwynn and Global Unichip now includes integrated hardware, optical engines, thermal monitoring and link-management capabilities.

The September 2026 financing announcement was also explicitly tied to manufacturing preparation, product validation and the expansion of its production ecosystem. Wiwynn's strategic investment adds a server manufacturer to Ayar's existing group of semiconductor backers.

Lightmatter has followed a different development path. Its work with Qualcomm demonstrated 1.6 Tbps per fiber, while collaborations with GUC, Cadence and Synopsys help customers integrate photonic technology into complex chip designs.

Lightmatter's Passage evaluation kits also give potential customers access to working hardware for testing. Avicena's 1 Tbps evaluation kits provide a useful comparison at an earlier stage of commercialization.

Across these companies, we are seeing optical development move from individual component demonstrations toward complete hardware, integration tools and manufacturing relationships.

What remains missing is large-scale production data: recurring shipment volumes, customer orders and the cost of operating these systems over time. Those figures will determine whether today's technical leaders can build substantial businesses.

Could a bigger semiconductor company buy the AI networking leader?

Acquisitions are a realistic outcome for AI networking startups, particularly optical companies whose technology could strengthen an established semiconductor manufacturer's product portfolio.

Marvell's acquisition of Celestial AI provides a useful example. The company completed the transaction in February 2026 after announcing approximately $3.25 billion in upfront consideration, with additional payments possible if certain milestones are achieved.

Celestial AI developed Photonic Fabric technology designed to connect computing and memory resources through optical links. Its technology operates in a market closely related to the opportunities pursued by Ayar Labs and Lightmatter.

The acquisition shows that a large chip supplier is willing to spend billions to gain access to optical interconnect technology before the market reaches maturity.

Ayar Labs and Lightmatter could attract similar interest from semiconductor companies seeking tighter control over optical connectivity. Xsight Labs could appeal to manufacturers interested in networking processors and switching silicon.

DriveNets would present a different proposition. Its reported $8.5 billion valuation, networking software and existing customer relationships would make an acquisition considerably more expensive.

An independent future is also possible. If optical companies become widely adopted suppliers to several chipmakers, remaining independent could allow them to serve a larger market than they would inside a single buyer's ecosystem.

What would make another AI networking startup overtake DriveNets?

Nexthop AI could overtake DriveNets commercially, while Ayar Labs or Lightmatter could become more important technology suppliers, but each company still needs to deliver specific results.

For Nexthop, the biggest missing piece is public evidence of sales. A disclosed multiyear contract with a major cloud provider, followed by recurring revenue and repeat orders, would make its $4.2 billion valuation much easier to assess. Nexthop has no public proof of sales, which is why we separate who has real customers from who only promises in our AI infrastructure report.

Ayar Labs needs to turn its integrated optical demonstrations into high-volume manufacturing. Its relationships with Wiwynn, Nvidia, AMD and other partners are valuable, but production shipments will be the decisive test.

Lightmatter needs to demonstrate that the extraordinary bandwidth density of its Passage technology translates into competitive system-level costs, power efficiency and reliable operation at scale.

DriveNets faces its own challenge. Its reported $1 billion in secured business includes telecommunications activities, so the company still needs to show how quickly AI-specific revenue is growing. A detailed breakdown of sales, new AI customers and profitability would give us a much clearer picture of its lead.

Competition from established manufacturers could also change the outcome. Nvidia remains the largest branded supplier in data center Ethernet switching, while Broadcom continues advancing its merchant switch silicon. Several startups depend on these manufacturers' chips, platforms or partnerships even while competing for some of the same customers.

So, which AI networking startup is ahead?

DriveNets is currently the leading independent AI networking startup overall, with Ayar Labs ahead in optical interconnect commercialization and Nexthop AI emerging as its strongest challenger in complete networking systems.

DriveNets has the clearest commercial record: more than $1 billion in secured business, positive cash flow since 2025, existing deployments with major network operators and a commercially deployed AI fabric connecting GPU clusters across two data centers.

Its lead is more convincing in commercial networking systems than across every underlying technology. Ayar Labs has made impressive progress toward manufacturing optical connections, while Lightmatter has demonstrated some of the highest bandwidth densities in the sector.

Nexthop AI has the capital and industry experience to become a major competitor, but we have not yet seen financial disclosures that put its actual business on the same level as DriveNets.

Meanwhile, Nvidia's latest quarterly revenue shows how far independent startups still have to go before they can challenge the established suppliers at market scale.

Our ranking is clear: DriveNets leads commercially, Ayar Labs leads the optical startup race, and Nexthop AI is the challenger most likely to change the overall ranking. The next major shift will come from large production contracts and recurring sales, rather than another record funding round.