Legend In distribution
Available through our vendor network at normal lead times — inquire for current pricing.
Supply Landscape
What we can source,
What we can source,
right now.
The current sourcing picture across our vendor network — which platforms are freely available, which are allocation-constrained, and where reserving early matters. Statuses reflect channel conditions, not a public price list: every engagement gets a firm quote with real lead times.
Last updated: July 8, 2026
Legend Constrained
Channel supply is tight. Allocation is reachable through vendor relationships — timing and volume matter.
Legend By reservation / Announced
Reserved-inventory programs and pre-launch platforms. We register early interest and secure allocation ahead of demand.
NVIDIA — Datacenter & AI
Status By reservation
NVIDIA GB200 NVL72
Rack-scale Blackwell platform for frontier AI training. Deployment-scale engagements only.
Status Constrained
NVIDIA HGX B200
Blackwell 8-GPU baseboard for AI training and inference clusters.
Status Constrained
NVIDIA B300 (Blackwell Ultra)
Blackwell Ultra generation for large-scale inference and training.
Status In distribution
NVIDIA HGX H200
Proven Hopper platform — strong availability and mature software stack.
Status In distribution
NVIDIA H100
Mature supply. Excellent value for established CUDA workloads.
Status Announced
NVIDIA Rubin platform
Next-generation architecture announced for 2026–2027. Early interest list open.
NVIDIA — Workstation & GeForce
Status In distribution
NVIDIA RTX PRO 6000 Blackwell
Flagship workstation GPU for visualization, simulation, and local AI.
Status Constrained
GeForce RTX 5090
Flagship consumer GPU — volume orders sourced through distribution.
Status In distribution
GeForce RTX 5080 / 5070 Ti / 5070
Full RTX 50-series stack for workstation fleets and system builds.
AMD — Datacenter & Client
Status Constrained
AMD Instinct MI355X
CDNA 4 accelerator — competitive option for large-scale inference.
Status In distribution
AMD Instinct MI325X
High-memory accelerator for inference-heavy deployments.
Status In distribution
AMD EPYC 9005 "Turin"
Current-generation server CPUs for AI head nodes and general compute.
Status In distribution
AMD Ryzen 9 9950X3D
Top desktop CPU for engineering and content-creation workstations.
Intel, Memory & Storage
Status In distribution
Intel Xeon 6
Granite Rapids server CPUs for storage and networking nodes.
Status In distribution
DDR5 RDIMM (Samsung / Micron / SK hynix)
Server memory at volume — spot and contract pricing.
Status In distribution
Enterprise NVMe (Samsung / Micron / WD / Solidigm)
U.2/E1.S datacenter SSDs for AI data pipelines and storage servers.
Status In distribution
Supermicro / ASUS / GIGABYTE server platforms
GPU servers, storage chassis, and rack infrastructure to complete the build.
Don't see the part you need? The list above is a snapshot of current demand, not a limit — we source across the full NVIDIA, AMD, Intel, Supermicro, ASUS, GIGABYTE, Micron, Samsung, and Western Digital catalogs. Tell us what you're standing up.
Supply Notes
Reading the market.
Occasional notes on where GPU and component supply is heading — written from inside the channel, not from press releases.
Blackwell supply is loosening — but not for everyone
A year after the Blackwell ramp began, hyperscaler commitments still absorb most of TSMC's advanced packaging output. What's changed: Hopper (H200) availability is genuinely good now, and B200 allocation is reachable for mid-size buyers who work through established channel relationships rather than open-market brokers. Our take: if your workload runs on Hopper today, the price-performance window is the best it has been; if you need Blackwell, plan allocation 1–2 quarters ahead.
The squeeze reaches the workstation
This month made it official: NVIDIA cut RTX 50-series supply to partners by roughly 20% with no desktop Super refresh coming in 2026, AMD told partners Radeon pricing rises 10–15% next quarter as GDDR6 spot prices tripled, and the RTX PRO 6000 Blackwell list price climbed to $13,250 — up 55% since launch. Memory is the common thread in every one of those sentences. Our take: for workstation fleets, buy before the Q3 increases land, and look hard at the RTX 5080 tier — it's the last part of the stack still selling near list.
Computex: the supply curve finally bends — slowly
Three things from Computex worth a buyer's attention: Vera Rubin is in full production with shipments starting this fall; Intel's Xeon 6+ "Clearwater Forest" is shipping day-one through the major OEMs and puts real competitive pressure back on server CPUs; and TrendForce now sees the CoWoS packaging gap narrowing from ~20% to ~10% by year-end. None of it makes hardware cheap — it makes 2027 look navigable. Our take: use the Xeon 6+ / EPYC rivalry for leverage on head nodes today, and pencil Rubin evaluations for early 2027 rather than fighting the fall ramp.
$375 for 32GB, and OEMs repricing daily
The cheapest 32GB DDR5 kit in the US just crossed $374 — the same capacity was $80–120 a year ago — and 64GB kits sit near three times their late-2025 price. On earnings calls the same week, Dell said it's repricing "it feels like, every day" and HPE flagged memory costs straining customer budgets. Our take: quote validity is the metric to watch now. When we send pricing, the memory line is the reason it expires in days — approve fast or expect a revision. Pre-buying RAM for known 2026 builds keeps beating waiting.
A $91B guide with China at zero
NVIDIA guided next quarter to $91B in revenue while assuming zero China datacenter business. Washington cleared H200 sales to a list of Chinese firms this month, but Beijing is holding imports at the border — so that demand stays out of the Western market for now. The uncomfortable read: if China reopens, an already tight allocation gets tighter overnight. Our take: the current window — H200 supply genuinely healthy, China sidelined — is the best sourcing environment Hopper buyers will see. We wouldn't bet on it lasting the year.
AMD just changed the second-source conversation
AMD's Q1: record $5.8B datacenter revenue, MI355X posting leadership MLPerf results, and Meta committing to up to 6GW of Instinct deployments. That last one matters — a hyperscaler betting production inference on Instinct is the validation mid-size buyers were waiting for. Our take: MI355X is now a real option for inference-heavy fleets, with strong memory capacity per dollar and shorter queues than Blackwell-class parts. Dual-sourcing isn't just leverage in the NVIDIA negotiation anymore; sometimes it's the faster path to racks that run.
Record margins upstream mean sticky prices downstream
TSMC posted a 66% gross margin, pointed record capex at AI capacity, and described wafer pricing as "value-based" — foundry-speak for going up. SK hynix followed with an all-time-high 72% operating margin quarter. When every layer of the stack is at record profitability with demand outrunning supply, nobody upstream blinks first. Our take: the cost structure in today's quotes isn't a spike, it's the new baseline. Budget 2026 projects at current pricing and treat any softening as upside, not as the plan.
The bottleneck moved: it's packaging now
News this week that NVIDIA is locking up TSMC's advanced-packaging expansion — reportedly more than half of the CoWoS capacity coming online through 2027 — as TSMC builds out packaging fabs in Arizona and Taiwan. Wafer starts were never the constraint this cycle; getting die onto interposers is. Our take: this is quietly good for NVIDIA-platform buyers, because the roadmap is supply-secured, and tough on anyone betting that alternative accelerators would loosen the market. Plan around NVIDIA lead times being predictable rather than short.
Allocation math: you ordered 100%, you're getting 60
Micron's blowout quarter came with the most honest sentence in the industry: key customers are receiving only half to two-thirds of the memory they order. That's what a real shortage looks like — not prices, fill rates. Meanwhile the first genuinely orderable Blackwell Ultra desk-side systems (DGX Station GB300 class) opened up through OEMs at quote-only pricing with late-summer lead times. Our take: treat memory line items like allocation, not procurement. Order early, over-spec the forecast, and keep a second qualified vendor for every module.
GTC: Rubin is real — buy Blackwell anyway
NVIDIA unveiled the Vera Rubin platform at GTC — Vera CPU, Rubin GPU, NVLink 6 — with partner systems landing in the second half of the year. Here's the trap in that keynote: waiting for Rubin means competing with hyperscalers for a brand-new ramp on day one. Blackwell Ultra is the "now" product with a defined supply curve. Our take: if you need compute in 2026, take Blackwell allocation and plan Rubin as a 2027 refresh. Announcements don't train models.
NVIDIA shipped a record $68B quarter and nothing got easier
Q4 FY26: $68.1B in revenue, $62.3B of it datacenter, gross margins around 75%. Read that margin number twice — it means NVIDIA has no incentive to cut prices, and relief has to come from supply growth, not discounts. Blackwell and Blackwell Ultra keep ramping through 2026, but demand is absorbing every unit. Our take: H100 and H200 are quietly becoming the value play. Hopper is available, mature, and understood — if your workload doesn't need Blackwell-class interconnect, the boring choice is the smart one right now.
Four companies pre-bought the year
Earnings week settled it: Amazon guided roughly $200B of 2026 capex, and with Microsoft, Alphabet, and Meta the big four are approaching a combined $700B — most of it AI infrastructure. That capital converts directly into GPU, HBM, and server allocation before the general market sees a single unit. This is why "just order it" stopped working. Our take: hyperscalers bought the front of the line; everyone else is buying relationships. Mid-size buyers who plan quarters ahead and work through the channel will still get hardware — the ones who shop spot will pay broker prices.
The memory market just reset the math on every build
TrendForce doubled its Q1 forecast this week: conventional DRAM contract prices are now expected up 90–95% quarter-over-quarter, with PC DDR5 up more than 100% and enterprise SSDs not far behind. We're seeing it in real time — module quotes that held for weeks last year are now good for days. The driver is structural: HBM for AI servers is eating the wafer supply that used to become commodity DRAM. Our take: if a build is funded, lock memory pricing now. Every week of waiting in this tape has been more expensive than the last.
Need firm pricing
and a real lead time?
Statuses tell you the landscape — a quote tells you your options. We typically respond within one business day.