AI inference is expanding semiconductor competition from GPUs to specialized chips and data center deployment: after SambaNova, Groq, and Cerebras, who will get the next ticket in?

AI Inference Is Expanding Semiconductor Competition from GPUs to Dedicated Chips and Data Center Deployment

The center of gravity in the AI computing market is shifting from training to inference. TechCrunch reported that a newly founded inference neocloud—General Compute—raised a $15 million seed round at a $60 million post-money valuation, betting on SambaNova’s inference chip SN50 and planning to provide compute services with air cooling, lower power consumption, and direct deployment in existing data centers. The financing size of this case is not large in itself, but it reveals a more important industry signal: value allocation across the AI supply chain is shifting from “who owns the strongest GPU” to “who can use the right chips to deliver compute to customers faster and more cheaply.”

This means that future competition will not only take place in the GPU market, but also across an entire chain that includes dedicated inference chips, neoclouds, data center power and cooling architectures, as well as advanced packaging and HBM supply. For the semiconductor industry, the demand structure for AI chips is changing; for the foundry market, the way advanced process and packaging capacity are used will also change; and for the supply chain, air cooling versus liquid cooling, data center retrofits, and power density requirements will all affect order structures for upstream equipment, materials, and packaging/testing firms.

In other words, this TechCrunch funding story is not just about a single startup; it is a window into how AI infrastructure spending is reshaping the semiconductor supply chain.

Background: From the Training Era to the Inference Era

Over the past two years, the main narrative in AI capex has centered on training clusters: the growing scale of large model parameters has driven extreme demand for NVIDIA GPUs, HBM, high-speed networking, and advanced packaging capacity. But as models gradually move into commercialization, inference is becoming the larger long-term market. Unlike training, inference places greater emphasis on cost per token, latency, throughput, and concurrency, rather than simply peak floating-point performance.This is exactly why dedicated inference chips have regained attention. In a TechCrunch report, it was noted that General Compute chose SambaNova instead of continuing to rely entirely on GPUs, partly because its architecture places greater emphasis on flexibility and memory capacity, making it well-suited to retaining context during inference. The report also pointed out that SambaNova’s new chip targets 600 to 700 tokens per second, compared with about 250 tokens for GPUs. Although such comparisons come from public statements by companies and founders and still require market validation, they send a clear signal: as AI moves from “generative models” to “agent-to-agent” workflows, speed, cost, and deployment model will determine the value of chips.

Industry Chain Analysis

Upstream: Changes in Demand Structure for Process Nodes, HBM, Materials, and Equipment

If inference chips continue to diversify, wafer manufacturing will face two kinds of pressure at once: one from the sustained demand for advanced process nodes from GPUs/AI accelerators, and another from the high-integration designs adopted by specialized chips to achieve power efficiency. Both types of chips rely on advanced process nodes, but their requirements for capacity, packaging, testing, and supply cadence are not the same.

• Foundry: TSMC remains the most critical advanced manufacturing platform. 3nm and 5nm are still important nodes for high-end AI chips and CPU/GPU combo designs; Samsung Foundry and Intel Foundry are competing for the next wave of AI-related foundry orders, especially projects that require faster customer customization and more complex packaging coordination.
• Advanced packaging: The bottleneck for AI chips is no longer only transistor density, but also chip-to-chip interconnects, HBM bandwidth, and packaging yield. If inference chips shift from a “single large GPU” to more distributed specialized compute modules, demand for advanced packaging will become more diversified, and the role of OSAT players such as ASE and Amkor will become more important.
• Equipment: Long-term demand for ASML, Applied Materials, Lam Research, and KLA remains tied to advanced-node and advanced-packaging capacity expansion. As long as AI drives investment in 3nm/2nm and subsequent packaging, equipment capex will not decline significantly.
• Materials: Higher-efficiency designs will continue to increase demand for high-end wafers, photoresists, specialty gases, and advanced packaging materials. Supply-chain resilience will matter more than price alone.

Midstream: Chip Design and Cloud Infrastructure Begin to Reshuffle Their Division of Labor

In the past, GPUs almost monopolized the narrative of AI training expansion, but the inference market is more likely to become layered:1. General-purpose GPUs will continue to handle the broadest software compatibility and training workloads; 2. ASIC/specialized inference chips will compete for specific models and specific latency-and-cost scenarios; 3. neocloud will package chip capabilities directly into services through compute leasing, data center consolidation, and customer onboarding.

This will have different impacts on NVIDIA, AMD, Google TPU, Amazon Trainium, and other paths. NVIDIA still has the strongest ecosystem, but inference-stage customers will care more about per-token cost, power consumption, deployment flexibility, and supply stability. AMD also has the opportunity to expand its presence among some cloud customers and enterprise AI deployments; while hyperscaler in-house chips, such as TPU and Trainium, if they can achieve lower TCO on inference workloads, market acceptance will continue to rise.

Downstream: data center deployment models are changing

The key feature of General Compute is not just that it uses SambaNova chips, but that it emphasizes air cooling, low power consumption, and deployment in existing data centers. This is very important. AI data centers have mostly been built around liquid cooling, high-density power delivery, and specially designed racks. If inference chips can lower the barriers to heat dissipation, they can significantly expand the range of deployable scenarios, and even give new value to some idle data centers, edge data centers, and legacy colo facilities.

For cloud infrastructure and data center operators, this means: the selling point in the future may not just be “more compute,” but “compute that is easier to bring online.”

Technology Impact

1. Inference optimization rather than pure peak compute

The value of inference chips lies not in single-point extreme performance, but in sustained throughput, memory hierarchy, and software stack compatibility. Architectures like SambaNova, which emphasize more flexible memory access and context retention, show that the industry is shifting from “brute-forcing compute” to “matching inference workloads.”

2. Advanced packaging remains a core battleground

Even if some inference chips do not pursue the absolute highest thermal design power, AI computing still depends heavily on chiplets, HBM, interposers, and high-density interconnects. In the long run, advanced packaging is not a byproduct of training chips, but part of AI chip competitiveness.

3. Air cooling / low power consumption may open new markets

If a certain type of AI chip can be deployed with less heat dissipation and lower power-delivery retrofitting, it will directly affect the speed of data center expansion. For customers, this means shorter delivery cycles; for the supply chain, it means incremental demand for liquid-cooling infrastructure may be diverted, but not disappear—instead, it will continue to be absorbed by higher-density training clusters and hyperscale cloud providers.

Supply Chain Impact

Who benefits?

• Specialized chip vendors: expanding inference scenarios open up design space beyond GPUs.- Specialized chip manufacturers: The expansion of inference scenarios broadens the design space beyond GPUs.
• Advanced packaging vendors: More AI servers and more heterogeneous chip combinations mean higher packaging complexity.
• Data center and colo operators: If next-generation inference chips become easier to deploy, existing data center assets can be repriced.
• Semiconductor equipment and materials suppliers: AI-driven investment in advanced process nodes and packaging continues to support long-term orders.

Who faces risks?

• GPU-only value chains: If inference workloads are diverted, growth in some GPU demand may fall below expectations set during the training peak.
• Data center solutions highly dependent on liquid cooling: If the market shifts toward low-power inference nodes, the economics of some heavy retrofit paths will deteriorate.
• ASICs lacking a software ecosystem: Specialized chips without enough developers and model adaptation can easily be marginalized in the inference market.

Competitive Landscape

From a competitive perspective, the AI computing market is becoming a “multi-tier market” rather than a single GPU monopoly.

• NVIDIA still remains the core price setter, especially with ecosystem advantages in training and high-end inference.
• Groq, Cerebras, SambaNova represent different forms of specialization: the former emphasizes low-latency inference, the latter emphasizes large-scale parallelism and system-level architectural innovation, while SambaNova is betting on more flexible memory and inference efficiency.
• In-house chips from cloud giants will continue to erode the standardized inference market, because they know their own workloads best.

Therefore, future market-share changes may not show up as “who replaces NVIDIA,” but more likely as: NVIDIA’s share of the total market remains very high, while specialized chips continue to increase their share in the inference segment.

Regional Implications

United States

The United States still holds the power to set the pace in AI chip design, cloud platforms, software ecosystems, and capital-market pricing. The emergence of neoclouds like General Compute shows that innovation in the U.S. market is happening not only at hyperscalers, but also at vertical service providers built around specific chip architectures.

Taiwan

Taiwan remains a key hub in advanced process nodes and packaging. As long as AI-related orders continue to concentrate in 3nm/2nm and advanced packaging, Taiwan will remain one of the most critical manufacturing nodes in the global supply chain.

South Korea

South Korea remains one of the core upstream players in DRAM/HBM and the memory supply chain for AI. The expansion of the inference market will not weaken the importance of HBM; it will only make high-bandwidth memory necessary for a broader range of AI chips.

Japan

Japan remains critical in materials, chemicals, and equipment components.### Japan

Japan remains critical in materials, chemicals, and equipment components. AI chip upgrades will continue to drive demand for high-purity materials and precision parts.

Europe

The biggest variable in Europe comes from the equipment segment, especially the EUV ecosystem around ASML. As long as advanced process nodes and packaging continue to advance, Europe’s strategic position in critical equipment will not change.

China and Southeast Asia

The Chinese market still has large-scale demand for AI inference deployment, but export restrictions will continue to affect access to high-end chips and equipment. Southeast Asia may continue to absorb some outsourced packaging and testing, assembly, and data center spillover demand, becoming a beneficiary region of supply chain diversification.

Investment Perspective

The reason capital markets are paying attention to this kind of company is not only because AI demand is growing, but because a new profit pool is emerging in AI infrastructure:

1. Compute services are shifting from “buying GPUs” to “selling outcomes”; 2. After specialized chips are combined with cloud deployment, a higher gross margin structure may emerge; 3. The valuation logic for data center assets is starting to shift from generic server rooms to scenario-specific AI infrastructure.

For investors, the truly important question is not “will the next Cerebras appear,” but rather: which architecture can establish enough software lock-in, supply chain stability, and customer stickiness in the inference era.

Long-Term Outlook

Next 3 years: Inference workloads will grow rapidly, GPUs will still dominate, but specialized inference chips and neoclouds will gain more experimental customers.

Next 5 years: The AI industry chain may become more clearly tiered, with training, inference, and edge deployment corresponding to different chip and data center architectures.

Next 10 years: If agentic AI becomes mainstream, inference cost will become the core competitive variable, and chip design and infrastructure will be rebuilt around “cost per result” rather than “peak performance.”

Conclusion

The partnership between General Compute and SambaNova is, on the surface, an early-stage financing round and the introduction of a new chip. In essence, however, it reflects a structural turning point in the AI industry chain: GPUs are still important, but they are no longer the only answer; advanced process nodes are still important, but packaging and deployment are becoming equally critical; data centers are still important, but the ability to deploy quickly within existing facilities is becoming the new threshold.For the semiconductor industry, the core judgment behind this trend is: the competition among AI chips has evolved from a single-dimensional compute race into a systems-level competition of “chip architecture + packaging + data center deployment + software ecosystem.” Whoever can drive inference costs lower, shorten deployment cycles, and make customer integration smoother is more likely to capture a larger share of the next round of AI infrastructure spending.

---

Source URL

• https://techcrunch.com/2026/05/28/has-the-hunt-for-ai-compute-uncovered-the-next-cerebras/