Should You Chase the Bounce? AI Infrastructure Passes the Baton as the Fed Hikes

Wall Street bounced the day after the Fed hiked a quarter point to a 3.75-4.00% target range and signaled another increase before year-end in the dot plot: the Dow rose 0.6%, the S&P 500 1.1%, the Nasdaq-100 1.7%, the Philadelphia Semiconductor Index 3.1% and the Russell 2000 about 0.6%. Presenter Jenny's point is that the move was less about a dovish Fed and more about the bond market having already priced the bad news. With the 2-year Treasury yield pinned near 4.679% and 10- and 30-year yields falling together, duration-sensitive growth stocks found breathing room.

The bond action was unusual: policy rate up, long yields down. Brent and WTI sliding below $100 helped by cooling energy-driven inflation pressure and giving the Fed and Treasury more control over the long end. The 10-year as a risk-free anchor anchors every valuation; while a 4-5% risk-free return demands a high equity risk premium, a softer long yield temporarily supports multiples. It does not mean the trend has turned, only that the most pessimistic path has been priced.

Breadth was solid and tech led. Nvidia added about 2.5% and AMD about 6%, less on a single catalyst than on the perception that AI infrastructure orders and roadmaps remain intact. Jenny argues the recent AI forum has moved debate away from bubble talk toward monetization: the last two years of pure training and brute-force compute accumulation are giving way to inference and agent use.

From training to inference: GPUs alone are not enough

The market translation is straightforward: GPUs remain essential but no longer sufficient. In task-oriented inference a model must repeatedly read parameters, process context and call tools, stressing memory bandwidth, network latency and data-movement efficiency. Rather than plowing the field with one processor, the whole chain — HBM, networking, cooling, data-center power and interconnect — must scale together. AI needs many flowers blooming at once.

Intel jumped 7-8% on two separate winds. First, SK Hynix is in talks to have memory made at Intel sites in the U.S., including Ohio; Reuters describes the discussions as early-stage but the order optionality was priced. Second, a broker lifted its target from $118 to $145 while the stock traded near $109, leaving upside. The inference narrative also helps the CPU story: GPUs stack brute-force compute while CPUs act as the conductor — reading web pages, running queries, orchestrating agent steps. The old 10:1 or 8:1 GPU-to-CPU ratio could compress toward 2:1 or even 1:1 for agent workloads, which is additive growth, not substitution.

The semiconductor bounce was not a one-stock affair. GlobalFoundries rose about 6%; it expanded a multi-year collaboration with Marvell to add capacity for optical transceivers, from pluggable modules to co-packaged optics and, eventually, co-packaged CPU optics. Marvell rose 4-5% on the news. At Nvidia, Jensen Huang, after last week's slowdown scare, framed demand resilience around sovereign AI and enterprise investment, saying chip volumes shipped next year could roughly double this year's. It is guidance in words rather than committed orders, but it injected confidence.

Pricing power test in neoclouds: Nebius hikes, bills rise

The cleanest micro story was Nebius. The company said it will hike selected pay-as-you-go list prices from October 1, its second increase in three months. Hourly lease rates for Nvidia H100, H200, B200 and B300 move up 17-21%, CPU-only services up 25% and certain memory services up 41%, all according to Reuters' coverage of the notice. The neocloud model is simple: burn capex first, build the AI factory, then rent it; revenue grows via more customers or higher price, ideally both.

The price anatomy matters: the hike hits pay-as-you-go small and mid-size users, while large hyperscale commitments from Microsoft or Meta keep discounts. That is a seller's market but carries two risks. If utilization falls or cancellations appear, hikes cannot stick and cuts follow. And even as rents rise, capex, new GPU prices, financing, power, networking and depreciation rise in parallel; revenue can climb while cash flow remains strained. The stock spiked close to 10% intraday to close up about 4%, still chopping in a range above the recent lows; a break above resistance could reopen the prior highs.

The same equation ran in reverse at CoreWeave, down over 4% on the day. Despite a similar business, the divergence is about the balance sheet: CoreWeave announced a $3 billion equity-plus-convertible-debt raise to fund more servers and halls, as reported by Reuters and CNBC that day. Even as contracted prices rise, the company cannot pause the investment cycle to defend competitive moats. So Nebius's pricing power was celebrated while CoreWeave's financing overhang dominated; the near-term risk is a false breakout, the medium-term catalyst is when operating leverage and unit cost actually inflect. In a high-rate regime, cash-flow quality and valuation discipline matter more than momentum alone.

Token economy and the power ceiling

The second half zooms from inside the data center to its outer bottleneck. A Seminare chart shown in the video estimates that the share of pre-training in combined Anthropic-plus-OpenAI compute falls from 67% in Q1 2024 to 7% in Q4 2026, while post-training and reinforcement learning rise to 55% and inference to 38%. The implication is that pre-training demands total compute, inference demands memory, network and data movement. A Citi compilation of OpenRouter data completes the picture: AI token consumption compounded at 31% month-over-month from January 2025 to August 2026, with August 2026 volume up 2,434% year-over-year. Tokens are where monetization happens, yet penetration remains low in the field. Most users still treat AI as a fancy search box, executing few agentic tasks; if platform plays like Apple's ecosystem quietly embed agents into devices, token demand could grow structurally. For investors the read-through is that long context and agentic reasoning lift demand beyond HBM into general DRAM, NAND and disk, plus networking and optical names such as Broadcom and Arista and even storage ETFs. A staged, trial-position approach beats chasing the bounce in fear-driven drawdowns.

The ultimate bottleneck is power. Amazon signed a long-term supply agreement with Generac; GNRC jumped 18% on the news. The deal locks in $2.4 billion of deliveries in 2027-28 within a framework that could reach $8 billion with follow-on purchases, plus warrants for up to 1.69 million Generac shares to Amazon. Generac, long known for residential and commercial backup generators, now pivots to AI data-center backup power, echoing backlogs stretching to 2028-29 at Bloom Energy and GE Vernova. Policy pushes the same way: the House passed the ratepayer-protection bill 417-3 to have data centers fund their own grid upgrades rather than socialize costs to households, a theme echoed previously by President Trump. The market chased the theme: Oklo rose about 13%, NuScale about 10% and a uranium ETF about 4%. But the picture is not rosy; Oklo is still down over 40% year-to-date, a pre-revenue developer facing Senate uncertainty and the need to convert policy-driven demand into bankable supply. Power, the bottom layer of the five-layer AI cake, remains the binding constraint and investable bottleneck for the next year or two.