A printed circuit board just pushed NVIDIA’s most powerful AI system into 2028 \u2014 at least according to one research firm. Semiconductor research firm SemiAnalysis reported on July 5\u20136, 2026 that NVIDIA’s Kyber NVL144 rack-scale AI system has been delayed, from its planned 2027 launch to 2028. CNBC cited the SemiAnalysis post. NVIDIA publicly disputed the report on Monday, with a spokesperson telling Bloomberg its “roadmap is intact.” <\/p>\n
NVIDIA (NASDAQ:NVDA<\/a>) shares actually rose more than 1% on Monday \u2014 helped by the company’s denial and a Goldman Sachs note calling the valuation compelling \u2014 leaving its market cap around $4.7 trillion. Overseas, PCB suppliers to NVIDIA came under pressure, with reports of sharp intraday declines in names such as Japan’s Ibiden and Hong Kong’s Kingboard Laminates Holdings. <\/p>\n What Kyber Actually Is <\/p>\n Kyber represents NVIDIA’s shift from selling chips to selling complete AI infrastructure. A single Kyber NVL144 cabinet houses 144 Rubin Ultra GPUs operating as one unified computing platform, targeting hyperscalers like Microsoft, Google, Meta, and Amazon. A larger sibling, the NVL576, designed to link eight racks via optical connections, is also likely delayed or limited to small initial volumes, according to SemiAnalysis. <\/p>\n Don’t wait: the analyst who called NVIDIA in 2010 just revealed his top 10 AI stocks. See the full list FREE now<\/a>. <\/p>\n The reported culprit is a component most retail investors have not heard of: the PCB midplane (which NVIDIA also calls the orthogonal backplane), a printed circuit board with up to 78 layers that connects electronic modules within the rack. SemiAnalysis said it “remains challenging from a manufacturability standpoint.” Stacking 78 layers with signal integrity, thermal tolerance, and yield at volume is a hard engineering problem, and SemiAnalysis frames it as the gating item for NVIDIA’s most ambitious system. <\/p>\n A Broader Pattern <\/p>\n The delay raises a broader question: whether NVIDIA’s annual cadence, built to stay ahead of AMD and custom silicon, is hitting physical manufacturing limits. Data Center revenue reached $75.246 billion (+92% YoY) in Q1 FY27, and CEO Jensen Huang described “the largest infrastructure expansion in human history”. The bottleneck is assembly. <\/p>\n The Competitive Opening <\/p>\n AMD (NASDAQ:AMD<\/a>) has the clearest window. Shares are up 157.77% year to date, and Goldman Sachs raised its AMD price target to $640 from $450 on July 6. CEO Lisa Su has said “customer engagement around MI450 Series and Helios is strengthening, with leading customer forecasts exceeding our initial expectations.” <\/p>\n Broadcom (NASDAQ:AVGO<\/a>) benefits through custom ASICs. Q2 AI semiconductor revenue was $10.8 billion, +143% YoY, with Q3 guided to $16.0 billion. Bloomberg Intelligence forecasts 27% CAGR for the custom ASIC market through 2033. <\/p>\n Google (NASDAQ:GOOGL<\/a>) is both a Rubin customer and a TPU competitor. Google Cloud revenue rose 63% with backlog nearly doubling to over $460 billion, and 2026 capex is guided up tp $190 billion. A Kyber slip strengthens the in-house silicon case. <\/p>\n The Bull Case <\/p>\n NVIDIA has not confirmed the delay. The affected product is a 2028 item; Blackwell and current Rubin systems remain in short supply. NVIDIA still commands approximately 81% of the AI chip market, and GuruFocus analysis flags NVDA as 45% undervalued relative to GF Value at current prices. Hyperscalers cannot swap vendors overnight. <\/p>\n Don’t wait: the analyst who called NVIDIA in 2010 just revealed his top 10 AI stocks. See the full list FREE now<\/a>. <\/p>\n<\/a> NVIDIA Blog \/ Press <\/p>\n