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: October 19, 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)
Now the dominant Blackwell SKU — large-scale GB300 server shipments reportedly began this month as partner availability broadens, with high-memory (288GB) configurations remaining in heavy demand.
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 By reservation
NVIDIA Vera Rubin platform
In full production and now shipping — first racks live at Microsoft Azure, with deliveries underway across eight cloud partners including AWS, Google Cloud, Oracle, and CoreWeave. NVIDIA pulled the Rubin CPX inference accelerator off the roadmap at GTC in late August, replacing it with a Groq-licensed LPU for decode workloads; broader channel allocation still expected in 2027.
NVIDIA — Workstation & GeForce
Status In distribution
NVIDIA RTX PRO 6000 Blackwell
Flagship workstation GPU for visualization, simulation, and local AI — NVIDIA's own marketplace listing hit $16,000 in September, up from $13,250 in June and 87% above original MSRP, as GDDR7 cost pressure keeps compounding; supply through distribution remains available.
Status Constrained
GeForce RTX 5090
Flagship consumer GPU — retail pricing has crossed $5,000 and US stock has largely vanished from major online retailers, with third-party marketplace listings running $6,500–$9,500 as AI builders buy cards in bulk; volume orders sourced through distribution.
Status Constrained
GeForce RTX 5080 / 5070 Ti / 5070
NVIDIA has reportedly cut GPU shipments to board partners by 15–20%; ASUS confirmed memory-driven production and restock constraints on the 5070 Ti and 5060 Ti 16GB after initial reports of discontinuation. No RTX 50 Super refresh is coming in 2026, and GDDR7 shortage keeps pushing pricing down the stack.
Status In distribution
NVIDIA RTX Spark (N1X)
Compact Grace-Blackwell "superchip" AI PC platform — Arm CPU cores paired with a Blackwell GPU and up to 128GB unified memory. The first wave of Windows systems — Lenovo, ASUS, Dell, HP, MSI, and Microsoft Surface — began shipping October 16; Acer and Gigabyte models follow in a later batch.
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 By reservation
AMD Instinct MI400 series (MI455X / MI430X)
CDNA 5, 2nm — 432GB HBM4 flagship (MI455X) anchors AMD's 72-GPU Helios rack. Helios entered full production this quarter, with first racks reaching hyperscale customers (Microsoft, Meta, OpenAI) and broader volume ramping into 2027.
Status In distribution
AMD EPYC 9005 "Turin"
Current-generation server CPUs for AI head nodes and general compute.
Status By reservation
AMD EPYC 9006 "Venice" (Zen 6)
Next-gen server CPUs on TSMC 2nm, up to 256 cores — ships as part of the Helios platform alongside MI455X, now in full production with broad OEM and cloud-provider rollouts beginning in Q4.
Status In distribution
AMD Ryzen 9 9950X3D
Top desktop CPU for engineering and content-creation workstations.
Status Constrained
AMD Radeon RX 9000 series (RDNA 4)
AMD raised official RX 9000 pricing 5–20% this quarter — the RX 9070 XT alone moved from roughly $830 to about $1,037 — citing rising GDDR6/GDDR7 costs; increases started in China and are rolling out globally.
Intel, Memory & Storage
Status In distribution
Intel Xeon 6+ "Clearwater Forest"
18A-built server CPUs, up to 288 cores — now in full production and shipping through major OEMs for storage, networking, and AI head nodes.
Status Announced
Intel Xeon 7 "Diamond Rapids"
Intel's Hot Chips deep-dive in August detailed the full package — 16 compute chiplets for up to 256 P-cores and 1.28GB of last-level cache on 18A-P — but reporting still points to a launch slip into 2027, putting it a generation behind AMD's EPYC Venice on timing.
Status Constrained
DDR5 RDIMM (Samsung / Micron / SK hynix)
SK hynix now says its DRAM, HBM, and NAND capacity for 2026 is essentially sold out, and Samsung raised 32GB DDR5 module pricing to $239 from $149 in September alone — a 60% jump. Samsung's preliminary Q3 results (released October 8) show operating profit near $80B — a quarterly record for a Korean company — with memory cited as the primary driver. The retail channel shows the same curve: the cheapest US 32GB DDR5-6000 kit crossed $450 in mid-October, more than double its price from a year earlier. Allocate now for anything not already committed.
Status Constrained
Enterprise NVMe (Samsung / Micron / SK hynix / SanDisk / Solidigm)
Micron's fiscal Q4 brought record NAND revenue of $14.1B, up 526% year-over-year, with data-center SSD revenue alone up roughly tenfold to near $10B and sequential NAND pricing up about 30%; Samsung, Micron, and SanDisk capacity stays presold well into 2027. CoreWeave's direct multi-year supply deal with Solidigm remains an early marker of AI clouds locking down flash allocation the way they lock down GPUs.
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.
TSMC's record quarter confirms the bottleneck hasn't moved
TSMC's Q3 report (released October 15) marked its sixth straight quarter of record profit, with high-performance computing — effectively AI accelerators — now 68% of revenue, and CEO C.C. Wei told analysts that demand still exceeds the company's supply capability even after this year's capacity additions. That's the clearest confirmation yet that 2026's constraint was never really about any one vendor's roadmap; it's foundry and packaging capacity that every AI chipmaker is competing for at once. Consumer memory keeps tracing the same curve from a different angle — the cheapest US 32GB DDR5-6000 kit crossed $450 in mid-October, more than double where it sat a year ago, confirming retail is still catching up to the contract-price moves distributors have been quoting for months. Our take: when the foundry itself is still behind demand after a record quarter, nothing downstream — GPUs, CPUs, or memory — gets structurally easier before 2027.
GPU allocation tightens again, and Micron's quarter shows why memory won't loosen
NVIDIA has reportedly cut GPU shipments to its board partners by 15–20%, and the RTX 5090 is the clearest symptom — retail pricing has crossed $5,000 with stock gone from major US retailers entirely, while AI system builders buying cards in bulk push marketplace resale as high as $9,500. Memory is the common thread: Micron's fiscal Q4 brought record NAND revenue of $14.1B, up 526% year-over-year, with data-center SSD revenue alone up roughly tenfold, and Samsung and SK hynix are projected to post a combined $130B+ in Q3 operating profit on the same dynamic. SK hynix told investors its DRAM, HBM, and NAND capacity for 2026 is now essentially sold out. Our take: this cycle has moved past the datacenter GPU line item — AIB shipment cuts and consumer-card scarcity confirm the memory shortage is now setting availability across the entire stack, workstation and gaming included, not just AI accelerators.
Earnings season confirms it: demand keeps beating capacity
NVIDIA's fiscal Q2 print (reported August 26) came in well ahead of its own guide and the company pointed to roughly $108B for the current quarter — a guide that keeps climbing even though TSMC's CoWoS packaging lines are fully booked through year-end and NVIDIA is estimated to hold the majority of that capacity. AMD's own August earnings told the same story from the supply side: Lisa Su said demand for both Instinct accelerators and EPYC server CPUs is running well above prior expectations, with unexpected demand now squeezing EPYC availability alongside GPUs. Samsung and SK hynix, meanwhile, reported finished-goods memory inventory at some of its lowest levels in years as HBM4 output for Rubin keeps pulling wafer capacity away from conventional DRAM. Our take: the list of bottlenecks hasn't changed — packaging, HBM, and now CPU lead times — it's just getting more crowded, so budget schedule slack across the whole stack, not just the GPU line item.
GTC's biggest inference chip wasn't NVIDIA's — and DRAM contracts lost their ceiling
At GTC in late August, NVIDIA pulled Rubin CPX off the roadmap entirely and handed its GDDR7-based inference role to the Groq 3 LPU, the first chip out of last year's $20B licensing deal — a reminder that platform roadmaps keep moving even after a product has been announced publicly. On the memory side, Q3 contract talks are getting worse for buyers, not better: SK hynix has reportedly dropped the price ceilings from its long-term supply agreements altogether, and Samsung is pushing for another double-digit increase, with Micron the last of the three still offering a floor-and-ceiling structure. Workstation and consumer GPU pricing moved in step — the RTX PRO 6000 Blackwell and the RTX 50 series both posted their steepest month-over-month increases of the generation in August, tracing back to the same GDDR7 and HBM allocation math. Our take: treat "no ceiling" contract language as the new normal rather than a negotiating tactic, and lock DRAM terms between cycles, not during them.
2027 memory capacity is already gone
Samsung, SK hynix, and Micron have reportedly sold out their entire 2027 DRAM and HBM capacity, and customers across the industry are currently being filled at only 60–70% of what they ordered. NAND tells the same story: Samsung, Micron, and SanDisk capacity is already presold, with Kioxia and SK hynix expected to finalize their own allocations by the end of this month. On the compute side, AMD's Helios platform — Instinct MI455X plus EPYC "Venice" — has moved from launch to full production, with first racks reaching hyperscale customers this quarter. Our take: this isn't a pricing story anymore, it's an inventory story — any 2026 build that still needs DRAM or NAND line items should lock them now, because the shelf for next year is already spoken for.
AMD puts dates on Zen 6 and MI400 — Rubin goes full production the same week
AMD's Advancing AI event (July 22–23) turned roadmap slides into shipping products: EPYC 9006 "Venice" (Zen 6, up to 256 cores on TSMC 2nm) reaches the channel in Q3, and the Instinct MI455X — 432GB of HBM4 in AMD's 72-GPU Helios rack — starts shipping to early customers this half with volume ramping into 2027. Days earlier, NVIDIA confirmed Vera Rubin is in full production, with hyperscaler deliveries beginning this fall. Our take: neither release changes near-term allocation math — Helios and Rubin racks are booked by hyperscalers first — but Venice gives buyers a credible upgrade path to plan budgets around, and Turin-generation pricing is worth locking now before attention (and wafer allocation) shifts to Zen 6.
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.