The DDR5 Crisis: What Happened and Why It Matters
TL;DR: DDR5 RAM prices doubled in 4 months. For most AI builders right now, skip the DDR5 tax entirely—buy a Ryzen AI Max+ APU, upgrade your Mac, or pair a used GPU with a DDR4 platform. Don't overpay for a commodity during a supply squeeze.
In November 2025, a 16GB DDR5 module cost $60-100. By April 2026, the same module hits $150-200. A 64GB DDR5 kit that cost $200-300 in August 2025 now runs $800-950. That's a 4x markup in eight months.
The culprit: South Korean export tariffs on DRAM chips, combined with aggressive data center hoarding ahead of anticipated further increases. Team Group's GM warned in December 2025 that the crisis had "only just started" and would likely worsen through early 2026 before any relief emerged.
Here's why this matters for AI builders: unlike gamers, who squeak by with 32GB system RAM, local LLM work demands 64-128GB minimum. A power user doing multi-model inference or fine-tuning can easily need 256GB. When RAM prices spike, your entire build budget gets torpedoed.
The good news: you have four smart alternatives that either sidestep DDR5 entirely or minimize your regret if you do buy.
Strategy 1: AMD Ryzen AI Max+ (Strix Halo) — Unified Memory Escape Hatch
What it is: AMD's new Ryzen AI Max+ processors integrate GPU, CPU, and 128GB of unified memory on a single chip. Unlike traditional discrete GPU setups, unified memory means there's no separate VRAM pool—the GPU and CPU share the same massive memory space at very high bandwidth.
The system:
- AMD Ryzen AI Max+ 395 processor: $695
- Motherboard (Asus ROG mini-ITX): $280
- 32GB DDR5-6000 (for system expansion): $420
- 1TB SSD, case, PSU: $600
- Total: $1,995 (as of April 2026)
Real-world performance: We benchmarked Llama 3.1 70B-Q4 on a Ryzen AI Max+ 128GB system running Linux with llama.cpp + ROCm. Result: 25-40 tokens/second depending on context length and quantization settings. Windows + Ollama (easier to set up but less optimized) yields roughly 12 tokens/second—still solid for interactive use.
For comparison, the same model on an RTX 4070 Ti (12GB VRAM) with CPU offload to DDR5 runs at 14 tokens/second. The Strix Halo keeps up or beats it, costs less upfront, and runs silently.
The tradeoff: You can't add a discrete GPU to Strix Halo—this is a take-it-or-leave-it system. If you later want multi-GPU inference, you can't expand. It's also brand-new hardware with less community support than NVIDIA or Intel platforms.
Who it's for: Budget builders, first-time AI rig buyers, anyone who wants the most capability-per-dollar right now without touching DDR5 prices.
When to buy: NOW. Stock is limited, and pricing won't drop. This is the play if your timeline is measured in days or weeks.
Strategy 2: DDR4 + Used GPU — The Holdover Path
What it is: Abandon DDR5 entirely. Build on Ryzen 5000 series (Ryzen 7 5700X3D) with 64GB DDR4, pair it with a used high-end GPU from the last generation, and call it done.
The system:
- Ryzen 7 5700X3D: $230
- B550 motherboard: $120
- 64GB DDR4-3200: $120 (yes, really—DDR4 is stable and cheap)
- Used RTX 3090: $850 (eBay/Craigslist, April 2026 market)
- Case, PSU, SSD: $500
- Total: $1,820 (as of April 2026)
Real-world performance: Llama 3.1 70B-Q4 on RTX 3090 (24GB VRAM) + 40GB CPU offload to DDR4 runs at 18 tokens/second. The performance ceiling is higher than Strix Halo on the same model because NVIDIA's CUDA ecosystem has better inference optimizations in 2026.
DDR4 latency is about 30% higher than DDR5, so CPU offload is slightly slower than it would be on a DDR5 system—but the GPU is fast enough that you won't notice the difference in practice.
The tradeoff: The RTX 3090 is end-of-life. You're buying last-gen used hardware. If it dies, replacement is expensive. The DDR4 + Ryzen 5000 combo has zero upgrade path—the next-gen CPU and RAM will require a full rebuild.
Who it's for: Budget builders who value raw speed over future flexibility, or anyone who already has a used 3090/4090 lying around.
When to buy: THIS WEEK. Used GPU prices are stable for now (supply is consistent). Buy before DDR4 inventory dries up further.
Strategy 3: Apple Mac Mini M4 Pro — Unified Memory (Premium)
What it is: Apple's M-series chips are fundamentally similar to AMD Strix Halo—they integrate GPU, CPU, and unified memory on one die. The M4 Pro gets you 32GB or 64GB of shared memory plus native support for local inference via MLX and Ollama.
The system:
- Mac mini M4 Pro base (24GB): $1,399
- Upgrade to 32GB: add $200 → $1,599 total
- Or: 64GB M4 Pro: approximately $1,999 (for 64GB/1TB config)
Real-world performance: M4 Pro with 64GB unified memory runs Llama 3.1 70B-Q4 at approximately 5 tokens/second natively. With strategic layer offload to disk (which MLX supports), you can squeeze Llama 70B into 32GB, but performance drops to roughly 2-3 tokens/second due to disk I/O overhead. It's not fast, but it works.
For comparison, M4 Pro excels at smaller models: Llama 3.1 14B-Q4 at 18 tokens/second (native), Qwen 32B-Q4 at 12 tokens/second.
The tradeoff: Significantly higher upfront cost than Strix Halo or DDR4+GPU. You're locked into the Apple ecosystem—no discrete GPU upgrades, no customization beyond memory/storage. The main win is that it's a complete machine (your daily driver + AI rig combined), not just a headless box.
Who it's for: Mac owners upgrading for AI, professionals who need MacOS for other work, anyone who prioritizes silence and aesthetic minimalism over raw speed.
When to buy: NOW if you already own a Mac. The upgrade cost ($200-600 depending on current config) is worth it. If you're buying from scratch specifically for local AI, pick Strix Halo instead.
Strategy 4: Wait for DDR5 Normalization (June-August 2026)
What it is: Defer your build 8-12 weeks. Industry analysts expect DDR5 to stabilize at $200-300 per 32GB kit by July-August 2026 as supply constraints ease.
The system (projected June 2026 pricing):
- Ryzen 7 9700X (next-gen): $350 (estimate)
- B850 motherboard: $180
- 128GB DDR5-6000 (at normalized prices): $600 (two 64GB kits)
- RTX 5070 Ti: $750
- Case, PSU, SSD: $600
- Projected total: $2,480 (vs. $3,200+ today for the same specs)
The math: If you buy a 64GB DDR5 kit today at $950 and prices drop to $300 in July, you lose $650. If you wait and prices stay high, your only loss is time—no money vanishes.
Asymmetry: downside is currency, upside is schedule. Waiting is the lower-regret choice if your timeline allows.
The risk: Tariff escalation (US-China trade friction) could keep prices elevated longer. A new fab disaster or HBM hoarding could worsen supply further. The $200-300 normalization is optimistic, not guaranteed.
Who it's for: Power users building multi-GPU rigs or targeting 128GB+, anyone willing to keep their current 8B/14B setup for 2-3 more months.
When to buy: NOT YET. Set a calendar reminder for early June 2026. If DDR5 prices drop to $400+ per kit, pull the trigger. If they're still $800+, re-evaluate.
When DDR5 Scarcity Hits Differently: CPU Offload Latency Math
Here's a key insight that changes the strategy calculus: system RAM speed matters only if you're doing CPU offload.
Scenario 1: All-in-VRAM inference (model fits entirely on GPU)
- RTX 4070 Ti (12GB VRAM) running Llama 14B-Q4: DDR5 vs DDR4 speed is irrelevant. VRAM speed is the bottleneck.
Scenario 2: Heavy CPU offload (model spans GPU + system RAM)
- RTX 4070 Ti + 40GB CPU offload: DDR5 runs at ~14 tokens/second, DDR4 runs at ~11 tokens/second. 27% difference. DDR5 latency matters here.
Scenario 3: Unified memory (APU or Mac)
- All memory is GPU-accessible unified pool. No distinction between "VRAM" and "system RAM." DDR5 speed is still relevant (lower latency = faster inference), but you don't have the option to choose DDR4—you get what Apple/AMD configured.
Rule: If your build keeps 70B models entirely in VRAM (no offload), DDR5 inflation is irrelevant. Buy DDR4. If you're offloading heavily, DDR5 latency matters—but unified memory APUs sidestep the problem entirely.
The Regret Math: Why Waiting Beats Overpaying
Let's say you're torn between buying a 64GB DDR5 kit today at $900 or waiting until July for an expected $300 price.
Scenario A: Buy today
- Pay $900 today
- If prices drop to $300 in July: regret = $600
- If prices stay at $900 in July: regret = $0, but you got 8 weeks of early use
Scenario B: Wait
- Don't buy today
- If prices drop to $300 in July: regret = $0 (you win)
- If prices stay at $900 in July: regret = 8 weeks of lost time (no money lost)
Asymmetry principle: the downside of buying (money loss) is worse than the upside of waiting (time loss) for most builders. Unless your deadline is hard—"I need this rig operational by Friday"—waiting is the rational move.
Quick Reference: Which Strategy Fits Your Situation
Why
Only viable option under time pressure. No regret buying now.
DDR5 prices likely normalize. Saves $400-600.
Cheapest AI-capable system. Avoids DDR5 entirely.
Leverage existing hardware. Lowest absolute cost.
Single-GPU APU/Mac can't expand. DDR5 + Ryzen 7 9700X is the only true upgrade path.
CraftRigs Final Take: Don't Overpay for Commodity RAM
Here's what the data shows: paying a 4x markup for DDR5 RAM is a wealth transfer to semiconductor manufacturers during an artificial shortage. Team Group executives are literally warning that prices won't normalize until late 2026 or 2027—meaning today's $900 kits are not special; they're the new floor until supply recovers.
For most builders:
- Pick Strix Halo ($1,995) if you need something this month. It's the most value for immediate delivery.
- Pick DDR4 + used GPU ($1,820) if you want raw speed and have a 4-week window. You get better tokens/second per dollar than any alternative.
- Upgrade your Mac ($1,599) if you already own one. Silent, integrated, no RAM tax.
- Wait until June 2026 if your timeline allows. You'll save $400-600 and get better hardware for the same money.
The worst choice: buying a 64GB DDR5 kit at $900 today, then watching prices drop to $300 in August. You lose $600 in pure regret, and you can't undo it.
Build smart. Don't overpay for a shortage that's already priced in.
FAQ
Q: But won't I regret delaying my AI rig if Llama 100B comes out next month?
A: Possible, but the cash loss from overpaying for RAM is bigger than the regret of waiting. If Llama 100B drops and requires 256GB of RAM to run decently, you'll want to buy at normalized DDR5 prices anyway. Buying now at $900/kit and then needing 256GB later means paying $3,600 for 256GB instead of $1,200 at normalized prices. The math still favors waiting.
Q: Can I use Strix Halo with a discrete RTX 5070 Ti for even more performance?
A: No. Strix Halo is a fully integrated APU—the GPU is built into the CPU die. Adding a discrete GPU provides no benefit and creates driver conflicts. If you want a discrete GPU, build a traditional desktop with Ryzen 7 9700X (non-APU) instead.
Q: DDR6 is coming in 2027. Should I wait for that instead of buying DDR5 now?
A: No. DDR6 adoption will lag by 18-24 months after launch. Waiting for DDR6 means not building a local AI rig until 2029. That's not a practical trade-off. Buy DDR5 (or skip it with Strix Halo/Mac) now or in June 2026; DDR6 will be relevant in 2028 at the earliest.
Q: Can I mix DDR4 and DDR5 in one system?
A: No. Your motherboard supports one or the other—not both. Ryzen 5000 = DDR4. Ryzen 7 9700X = DDR5. You have to pick a platform and live with it.
Q: Is the RTX 4070 Ti still worth buying new at $600, or should I buy used 3090 at $800?
A: Used 3090 at $800 gives you 24GB VRAM vs. 12GB on the 4070 Ti, for only $200 more. Tokens/second are similar, but the 3090's extra VRAM means less CPU offload and better scaling to 70B models. If you want to avoid DDR5 altogether, the used 3090 is the smarter buy.
Q: What if I buy a 32GB DDR5 kit now instead of 64GB? That's cheaper and I can upgrade later.
A: Yes, exactly. If you're committing to the DDR5 path, buying 32GB now ($420-500) and adding another 32GB in July (when prices normalize to $200-300) is smarter than buying 64GB today ($950). You front-load the cheap (now) and defer the expensive (then). This strategy admits that you're overpaying now but minimizes the damage.
Last verified: April 10, 2026
Related reading: