As artificial intelligence infrastructure funding balloons to unprecedented levels, concerns over an "AI credit bubble" and rising long-term Treasury yields have been weighing on tech stocks tied to the AI computing theme. Yet, a fresh research note from BNP Paribas analysts argues that the Hot Chips conference—often dubbed the "Olympics of semiconductors"—has delivered a critical positive signal for the AI compute supply chain: innovation is accelerating in tandem across the four core pillars of AI computing, memory, networking, and advanced packaging, from OpenAI's custom inference chip "Jalapeño" and NVIDIA's BlueField-4 network architecture to AMD's optical interconnect in MI500 and new HBM advancements from Samsung Electronics and SK Hynix.

NVIDIA's just-released fiscal Q2 2027 results showed revenue surging 106% year-over-year to $96.221 billion, with adjusted EPS of $2.22 beating already-lofty analyst expectations by more than 5%. Data center revenue alone jumped 117% to $89 billion. More striking, management led by Jensen Huang projected fiscal 2028 revenue growth of roughly 70%, significantly above the pre-earnings Wall Street consensus of about 50% based on $1 trillion in order visibility for Blackwell and Rubin. The company guided Q3 revenue to $108 billion, plus or minus 2%, which combined with the Hot Chips signals provides further proof that global AI compute demand remains in a massive expansion cycle that is far from peaking.

The BNP Paribas team, led by Karl Ackerman, points out that the unstoppable penetration of Agentic AI across industries is shifting system bottlenecks away from raw GPU compute throughput toward inference-stage memory bandwidth, KV cache, low-latency interconnect, real-time task orchestration, and data processing. This means the AI compute investment theme is expanding from a "GPU and TPU solo act" into a complete and broad AI infrastructure stack encompassing custom XPUs/AI ASICs, high-bandwidth memory, optical interconnects and NICs, plus DPUs and CPUs. Major Wall Street financial giants including Goldman Sachs, Morgan Stanley, and Bank of America all recently echoed that the AI supercycle is far from over—it is transitioning from an "AI chip buying frenzy" into a second phase of "building AI factories at scale." The next wave of outsized alpha returns will no longer belong solely to the strongest leaders in AI GPU/AI ASIC space, but will systematically spread to high-performance data center CPUs, DRAM/NAND/HBM storage, AI PCBs, liquid cooling systems, optical interconnect/communication systems, ABF substrates/glass substrates, MLCCs, data-center-grade electronic fabrics, and broad wafer foundry beyond advanced process nodes across the entire "AI factory" full-stack infrastructure layer.

With NVIDIA again delivering blowout results and an explosive outlook, AI computing trading hotspots are likely to accelerate beyond NVIDIA's GPU clusters to encompass the entire AI compute supply chain—including HBM/DRAM/NAND, CoWoS/3D advanced packaging, data center CPUs, high-performance network infrastructure, optical interconnects, and data center power chain infrastructure—sparking a fresh, chain-wide "main upswing" super rally.

AI Credit Risk Can't Suppress the Inference Revolution: OpenAI's Custom Chip Debuts, NVIDIA Accelerates 'Scale In,' and the AI Infrastructure Stack Enters a Full-Spectrum Bull Market

The BNP Paribas research team highlights OpenAI's Jalapeño and NVIDIA's network architecture as key takeaways from Hot Chips. Ackerman's team detailed critical outputs and research conclusions from the conference, noting that AI agents focused on fully automated agentic workflows are the catalyst driving innovation, further underscoring the enormous potential value of AI infrastructure as an investable theme.

The analyst team led by Karl Ackerman outlined the following key points. First, they state that OpenAI's custom-built inference chip (an AI ASIC approach similar to TPU) named "Jalapeño," co-developed with Broadcom, is a major positive for Broadcom Inc (NASDAQ: AVGO) and Celestica Inc (NYSE: CLS). Second, Google parent Alphabet Inc (NASDAQ: GOOGL) has made a major commitment to launch two cutting-edge advanced-process chips per year for each product generation. Third, Advanced Micro Devices Inc (NASDAQ: AMD) plans to integrate high-speed optical components—leveraging silicon photonics via the CPO technology route—into the scale-up interconnect of its upcoming MI500 AI chip.

Fourth, the analysts note that NVIDIA's AI network architecture "Scale In" is powered by its key accelerated infrastructure chip, the BlueField-4 data processing unit (DPU). Fifth, Facebook parent Meta Platforms Inc (NASDAQ: META) is adopting Broadcom's Tomahawk Ultra for scale-up purposes. Sixth, South Korean memory giant Samsung Electronics is innovating in high-bandwidth memory (HBM) base dies, while fellow South Korean memory leader SK Hynix Inc (OTC: SKHY.US) is willing to adopt Intel's advanced packaging technology—embedded multi-die interconnect bridge (EMIB)—for its HBM product line's advanced packaging system. Finally, the analysts say AMD's Helios has a significantly higher network interface card (NIC) attachment/connection rate per server than NVIDIA's Vera Rubin, which is a major positive for the data center optical interconnect segment—specifically the high-speed optical transceiver ecosystem.

The analysts also point out that the proliferation of AI inference-driven super applications is accelerating innovation at the memory layer. Ackerman and his team state: "The current memory hierarchy is primarily oriented toward training workloads, but inference is inherently constrained by overall memory performance. Moreover, as context length and user numbers grow, KV cache demand continues to rise. Therefore, we believe SRAM, 3D-DRAM, Vertical HBM, and high-bandwidth flash (HBF) are all seeking to meet the memory bandwidth, latency, and robust power requirements of being closer to—or directly integrated onto—the core AI compute processors."

Additionally, the analysts note they are seeing an increasing number of custom, inference-optimized AI accelerator system iterations, including Google TPU, Meta MTIA, Microsoft Maia, OpenAI Jalapeño, Cerebras WSE, and Groq LPX. They add: "We attribute the proliferation of new XPUs/AI ASICs to the diversity of AI inference workloads and the highly attractive token economics of compute infrastructure." Ackerman's team also says agentic CPUs are taking center stage. As workloads shift toward multi-step workflows, the system bottleneck transforms into real-time task orchestration and data processing. Different CPU approaches are emerging: NVIDIA Vera targets the highest level of AI single-thread performance; AMD EPYC Venice takes a portfolio-based approach to serve diverse workloads; and Intel's exclusive Diamond Rapids differentiates through its fan-out fabric design.

The BNP Paribas team has raised its price target for Broadcom to $675, up from $640, implying a potential upside of up to 90% over the next 12 months. For Celestica, the target is now $500, up from $450, implying a potential upside of 60%. For the "AI chip super king" NVIDIA, the target stands at $285, implying nearly 40% upside.

70% Growth Guidance Pierces the 'AI Compute Peak' Theory: Is Vera Rubin Igniting a New AI Compute Supply Chain Bull Run?

NVIDIA management's Q3 revenue guidance midpoint reached $108 billion, while the company unusually projected approximately 70% fiscal 2028 revenue growth, well above the pre-earnings market expectation of roughly 44%, emphasizing that this figure is constrained by supply capacity rather than demand. The signals from Hot Chips, combined with NVIDIA's strong results and outlook, are accelerating the global AI compute infrastructure investment theme into a "second-round industry boom" driven jointly by Agentic AI and the massive Vera Rubin upgrade cycle.

Vera Rubin has already begun volume shipments, expected to contribute about 20% of data center revenue this quarter. This is compounded by Anthropic's $45 billion six-year compute deal with Nscale—which utilizes NVIDIA's latest Vera Rubin compute—and Amazon Web Services' (AWS) plan to deploy an additional 2 million NVIDIA GPUs in 2027-2028. These developments show that demand has expanded from traditional hyperscale cloud providers to frontier model labs, NeoClouds, sovereign AI, and a broader global enterprise customer base, with order visibility showing no signs of "compute peaking."

Hot Chips further reinforces that this AI supercycle is upgrading from a "GPU solo act" to full-stack AI compute infrastructure expansion: OpenAI's Jalapeño combines custom inference ASICs with Broadcom networking technology and Celestica rack systems; AMD is preparing to introduce optical technology in the MI500 scale-up interconnect; NVIDIA is strengthening Scale In high-performance networking with the BlueField-4 DPU; Meta is adopting Tomahawk Ultra; and Samsung Electronics, SK Hynix, and Intel's EMIB advanced packaging are all competing for HBM upgrade dividends. The underlying logic of these frontier technology iterations is that agentic inference, planning, and multi-round tool invocation generate far more compute demand than traditional Q&A. The system bottleneck is shifting from single-GPU compute power to KV cache, memory bandwidth, low-latency networking, optical interconnect, and CPU real-time orchestration. Therefore, OpenAI's custom chip is not a signal of weakening AI compute demand but rather a result of no single vendor being able to cover all workloads at optimal cost. It presents localized competition for NVIDIA's share while simultaneously expanding the total addressable market for AI compute supply chain leaders like Broadcom and Celestica, as well as HBM, optical modules/communications, AI PCBs, liquid cooling, data center CPUs, and 3D advanced packaging.

The Anthropic-Nscale deal translates this shift from "GPU solo act" to full-stack infrastructure expansion into long-term capital commitments. The six-year compute lease locks in approximately 460 megawatts of capacity in West Virginia, with Nscale deploying Vera Rubin systems, replacing campus space freed up after Microsoft withdrew its letter of intent. A single model lab signing a long-term contract for nearly half a gigawatt of compute massively strengthens order certainty for the entire rack-scale compute clusters led by Vera Rubin, data center power, and NVIDIA-led high-performance network infrastructure equipment.

Furthermore, top Silicon Valley venture capital firm a16z has recently observed that hourly rental prices for NVIDIA B200 AI GPU components are rising against the trend, models are becoming professionally differentiated, and consumer willingness to pay for AI applications is increasing. This largely implies that compute demand is transforming from one-time training capital expenditure into recurring inference operational expenditure, making the AI compute infrastructure cycle more durable.

Wall Street analysts' average price target for NVIDIA implies up to 45% potential upside. Currently, the most bullish public target on Wall Street comes from Baird senior analyst Tristan Gerra at $500 with an "Outperform" rating, implying a potential stock gain of approximately 132.9% and a market cap of roughly $12.2 trillion—an increase of about $7 trillion from current levels. This market cap projection assumes largely unchanged share count. Baird's aggressive bullish thesis on NVIDIA rests on four core pillars: First, NVIDIA is further expanding share in inference computing and hyperscale cloud customers, with agentic AI requiring significantly more inference, planning, and tool invocation per task than traditional chatbots. Second, Vera Rubin's adoption speed and deployment scale are expected to exceed Blackwell, continuously driving up the value of GPUs, NVLink, networking, and full rack systems. Third, the standalone Vera CPU could penetrate the traditional x86 server market thanks to bandwidth and efficiency advantages, which Baird estimates could open up an incremental market of approximately $200 billion for NVIDIA. Fourth, global AI infrastructure spending could surpass $1 trillion in 2027 and reach $3-4 trillion annually by 2030, with NVIDIA's CUDA ecosystem, chip-network-system full-stack capability, and annual product cadence forming its most difficult-to-replicate moat.