Digital Colliers Daily Briefing — July 6, 2026
The AI infrastructure stack is showing simultaneous stress at three layers today: silicon roadmaps, memory financing, and macro sentiment. Nvidia has slipped its next rack-scale architecture by more than a year, SK Hynix is tapping US markets for one of the largest tech equity raises on record, and a draft Treasury report is quietly drawing parallels between the current AI cycle and the dotcom bust. Together, the three data points sketch a market that is still spending aggressively while its underlying schedule and confidence assumptions wobble.
1. Nvidia pushes Kyber NVL144 to 2028, scraps NVL72x2 stopgap
What happened. According to a SemiAnalysis note surfaced by CNBC's Anniek Bao, Nvidia has delayed its Kyber NVL144 rack system — the platform designed to house the 2027 Rubin Ultra generation — by at least 12 months, pushing volume availability into 2028. The cited cause is PCB manufacturing issues tied to the Kyber design. Nvidia has also cancelled the NVL72x2 architecture that had been positioned as an intermediate doubling of the current NVL72 platform.
Why it matters. Rack-scale systems, not individual GPUs, have become Nvidia's unit of competition against custom silicon from Google, AWS, and Meta. A year-plus slip to Kyber removes the marquee 2027 platform hyperscalers were sizing power, cooling, and data-center shells around, and the NVL72x2 cancellation eliminates the bridge product meant to absorb demand in the interim. Compute roadmaps built on Rubin Ultra timing — including customer capex disclosures made earlier this year — now need to be reworked against a longer Blackwell Ultra and Rubin runway.
Who is affected. Every hyperscaler with 2027–2028 buildout commitments: Microsoft, Google, Meta, Oracle, AWS, and the growing tier of neocloud operators like CoreWeave and Crusoe. PCB and substrate suppliers keyed to Kyber lose a year of shipment volume, while incumbents on the Blackwell and Rubin baseline platforms gain extended demand. Custom-silicon programs at Google (TPU), AWS (Trainium), and Meta (MTIA) get more runway to close the systems-integration gap Nvidia had been widening.
What to watch next. Any Nvidia commentary at its next earnings call reconciling the SemiAnalysis reporting; revised Rubin Ultra sampling dates; and whether hyperscalers publicly re-baseline 2027 capex or quietly reallocate toward custom accelerators and additional Blackwell/Rubin racks.
Sources:
2. SK Hynix launches ~$28B Nasdaq share sale to fund HBM expansion
What happened. SK Hynix on Monday launched a US share sale on Nasdaq, offering 17.79 million new shares to raise roughly 43 trillion won, or about $28.07 billion, per Reuters. Final pricing is due Thursday, with trading set to begin Friday. The offering runs alongside the company's existing Korea Exchange listing.
Why it matters. At $28 billion, this ranks among the largest equity raises the technology sector has ever executed and is the clearest signal yet that HBM capacity — not logic wafers — is the binding constraint on AI system output. SK Hynix supplies the majority of HBM3E consumed by Nvidia's Blackwell and Blackwell Ultra platforms and has been the pace-setter on HBM4 qualification. Financing at this scale suggests the company intends to widen its capacity lead over Samsung and Micron rather than defend a narrower one, and it validates a multi-year demand curve well beyond current order books.
Who is affected. Nvidia and AMD as primary HBM customers gain visibility that supply expansion is funded through the decade. Samsung and Micron face a competitor with a fresh war chest at exactly the moment HBM4 qualification determines 2027 share. US equity investors get direct exposure to a name previously accessible mainly via ADRs and KOSPI. For Seoul, a Korean national champion raising dollars on Nasdaq is a notable capital-markets shift with political overtones.
What to watch next. Thursday's final pricing and the implied valuation versus the KOSPI line; disclosures on how proceeds split between M15X, Cheongju, and any US fab footprint; and whether the raise coincides with tightened HBM4 supply agreements with Nvidia. The timing also sits awkwardly against Event 3 — a large equity offering into a market a government report is quietly comparing to 1999.
Sources:
3. Draft Treasury report draws dotcom parallels for AI market
What happened. A draft US Treasury Department report obtained by NOTUS's Eric Katz warns of risks in the AI market and likens key aspects of the current cycle to the dotcom crash of the early 2000s. The document sits in contrast to the Trump administration's public posture, which has been consistently bullish on AI investment and deployment.
Why it matters. Treasury analyses of asset-price risk rarely leak without consequence, and a formal comparison to the dotcom bust — even in draft form — establishes a reference point that regulators, congressional staff, and institutional allocators will now cite. The concern is well-founded on scale alone: hyperscaler AI capex is running at multiples of late-1990s telecom infrastructure spending, and revenue attribution for that spending remains thin outside a handful of vendors.
Who is affected. Publicly traded AI infrastructure names — Nvidia, Broadcom, Oracle, the hyperscalers, and the neocloud cohort — face a new narrative headwind independent of fundamentals. Private AI companies raising at 2025-vintage multiples may find LP diligence sharpening. Financial regulators at the SEC, OCC, and Fed gain implicit cover to ask harder questions about bank exposure to AI-linked lending, GPU-collateralized debt, and concentrated data-center financings.
What to watch next. Whether the final report retains the dotcom framing or softens it under White House review; any congressional hearings citing the document; and whether credit-rating agencies begin adjusting outlooks on AI-heavy issuers. Investor reaction to Friday's SK Hynix debut will be an early real-time read on whether the warning meaningfully shifts sentiment.
Sources:
The three stories rhyme uncomfortably. Nvidia's Kyber slip shows that even the vendor best positioned to execute is hitting manufacturing limits at the systems layer; SK Hynix's $28 billion raise shows that the capital required to keep pace has moved beyond what conventional balance sheets can absorb; and Treasury's draft suggests that at least some inside government are beginning to price in the possibility that the demand assumptions justifying all of it may not hold. None of these individually derails the AI buildout, but read together they mark the point in the cycle where schedule risk, financing risk, and sentiment risk are being discussed in the same week.