AI Chips Beyond Cloud Need 18 Months, HCLTech and Omdia Say

An 18-month estimate for AI beyond the cloud

HCLTech has put a number on a question that hangs over the AI hardware boom: how long before AI silicon moves past hyperscale data centers and into broad commercial use? The company's answer is about a year and a half. It says the industry is still building the integrated semiconductor platforms needed to support AI at scale outside the cloud.1 The estimate comes from a white paper produced with research firm Omdia and released on November 3.1 The paper is titled AI Semiconductor Platforms: Powering the Next Wave of Global Growth – An Asia Pacific Perspective, and much of it is about Taiwan's place in AI supply chains.31

Public coverage gives little detail on how the 18-month figure was reached. The direction is clearer, though. According to the paper, AI deployment is spreading from hyperscale facilities into enterprise infrastructure, telecom networks, factories, vehicles, PCs, smartphones and other edge devices. That widens chip demand to include specialty and mature-node parts used for sensing, power management and control, not only leading-edge processors.35

I read the forecast as a statement about engineering effort more than chip availability. The silicon already exists. The wait is for platforms that can be packaged, cooled, programmed and deployed outside the controlled conditions of a hyperscaler's data center.

Platforms, not chips

The paper's central argument, repeated in every write-up of it, is that AI semiconductor competition has moved away from raw chip speed. What matters now is how well compute, memory, connectivity, packaging and software work together as a system.32 Ameer Saithu, who heads HCLTech's semiconductor business, said the winners will be judged by workload results rather than benchmark wins, and that execution is the real separator.31

To show how complex this has become, the paper cites Nvidia's GB200 NVL72, which puts 72 GPUs and 36 CPUs in one liquid-cooled rack. It argues that systems like this require compute, connectivity, power and cooling to be designed together.32 The paper also points to advanced packaging, especially TSMC's CoWoS, as a key step for combining logic with high-bandwidth memory.32

The Taiwan figures explain why the island is at the center of the analysis. Taiwan-headquartered foundries hold more than half of the global foundry services market. Taiwanese original design manufacturers make about 90% of the world's server motherboards and 80% of server systems. Omdia puts TSMC's share of leading-edge production above 90%.32 ASE Technology Holding has roughly 40% of the outsourced assembly and test market.31 The paper also names Singapore, Malaysia, Vietnam and China as part of a more spread-out Asian network, while still treating Taiwan as the connecting hub.32

HCLTech has a business interest here. The company has been moving its engineering services toward silicon, and HFS Research describes a "chip-to-cloud" strategy that includes outsourced assembly and test capability and a semiconductor lab in Bangalore.4 In July it announced plans to invest up to ₹3,500 crore in AI data centers scalable to 50MW.8 A white paper saying platform integration is the bottleneck also argues for the kind of integration work HCLTech sells. That doesn't make the thesis wrong, but it should be read with that in mind.

Nvidia's roadmap supports the thesis

Nvidia's latest products match the paper's view that platforms matter more than single chips. Vera Rubin is now described as a seven-chip platform. It combines the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch and a Groq 3 LPU.13 Nvidia now sells a whole rack, not a single GPU.

Coverage of Rubin's timeline is inconsistent. Some reports say Nvidia declared full production at CES in January, began ramping in May, and planned to ship in the fall.17 Another report dates the full-production announcement to October 21, with Nvidia saying production shipments would start this fall.14 On September 30, Nvidia said CoreWeave was offering Vera Rubin NVL72 to early-access customers, and Blackwell remained the company's volume revenue product into its first fiscal quarter of 2027.13 My reading is that "full production" has meant different things at different times. Rubin is only now reaching customers in real volume, which fits HCLTech's view that deployment trails announcements.

The headline specifications are large. Each Rubin GPU is rated at 50 petaflops of FP4 inference with 288GB of HBM4.12 Nvidia has said the platform can cut inference token costs by up to 10x compared with Blackwell.12 The October announcement used a different measure: 30x higher throughput per megawatt for agentic workloads.14 One analysis cautioned that neither figure proves a general tenfold gain.14

Nvidia is also making a separate push outside the data center. At GTC Berlin it announced that DGX Spark now supports eight-unit clusters, with inference speedups of up to 40% on some RTX and DGX systems for local AI development.16 In October, Nvidia and Microsoft introduced RTX Spark, a Windows AI PC platform with up to 128GB of unified memory.11 These are early versions of the beyond-cloud market the white paper describes, and they suggest Nvidia expects to compete there too.

Inference changes the hardware

Inference is now driving hardware design. IEEE Spectrum quoted Jensen Huang as saying that inference is all chief information officers want to discuss, and calling this moment an inflection point.20 Nvidia's answer is to split the work across two chips. Rubin GPUs handle the compute-heavy prefill phase, and the Groq 3 LPU, with its large on-chip SRAM, handles memory-heavy decoding.20 Nvidia got that technology by licensing Groq's IP and hiring its staff in late 2025, not by buying the company.1320

The split makes the paper's integration argument concrete. Splitting inference across different chips raises the importance of interconnects, orchestration software and rack design. Those are the system-level layers the paper says will decide the winners.

Google's TPUs make inference costs the battleground

The strongest challenge to Nvidia's pricing power is coming from custom silicon, and Google's TPUs lead that effort. In September, SemiAnalysis benchmarks reportedly put Ironwood's cost at $0.181 per million tokens at 100 tokens per second per user. The same tests put Nvidia's B200 at $0.222 and B300 at $0.276, which makes Ironwood about 18% cheaper than B200.21 SemiAnalysis has separately estimated Ironwood's total cost of ownership at about 44% below an Nvidia GB200 server.27 Google Cloud chief Thomas Kurian has said servers running Google's own silicon pay back in roughly half the time of its other AI servers.27

The cost data is less tidy than those numbers suggest. One analysis said Google had not published an Ironwood chip-hour price months after general availability, and that the only firm figure was a SemiAnalysis estimate of about $1.60 per hour for Anthropic.23 Other reports cite list prices of $12.00 per chip-hour on demand and $5.40 on a three-year commitment.2124 At that list price, one detailed analysis concluded that Ironwood costs about the same as a B200 on Google Cloud.24

SemiAnalysis's newest data shows how quickly the lead can change. On Kimi K3, Ironwood was competitive with Nvidia's GB300 at FP4 when it launched. After four Nvidia software refreshes, GB300 now leads by 1.5x to 2.6x, and Nvidia's VR200 is shown delivering 4.3x to 5.6x Ironwood's tokens per dollar.30 Ironwood still came out as the cheapest FP8 option for serving GLM-5.2.30 Taken together, custom silicon wins on some workloads, while Nvidia's software and next-generation hardware keep moving the comparison.

Google is also changing how it sells TPUs. It is using partners to put more of them into the market,27 and distribution through Oracle reportedly expanded in late October.21 Google Cloud says its TPU 8i offers up to 80% better inference performance per dollar than Ironwood.26 Lock-in is the main drawback. One analysis estimated two to eight weeks of engineering work to move to TPUs, along with the ongoing cost of maintaining a separate software stack.22

What the 18-month window means

Taken together, the coverage makes HCLTech's estimate look reasonable, if a little optimistic. In the cloud, the cost per token is falling fast because Nvidia and Google are competing on rack-scale systems and custom silicon. Moving that capability to factories, vehicles and enterprise sites, as the white paper anticipates,35 depends on the packaging, cooling and software integration the paper describes. Omdia's own 2Q26 research points to bottlenecks in wafers, HBM and CoWoS.38 Those limits will affect how fast AI hardware spreads beyond hyperscalers.

I expect the next 18 months to look like this. Cloud inference prices keep falling, edge and enterprise AI arrive mainly as platform products such as Nvidia's Spark systems, and Taiwan's packaging and server capacity decides the pace more than any one chip announcement does.