← The AI Hype Audit — all 645 verdicts
PARTLY
Partly true: the $260 dual-socket X99 board really has six PCIe slots and six old Tesla P100s really add up to 96GB, but 'same VRAM as a $25,000 Blackwell workstation' compares 2016 cards to 2026 ones
The claimSix GPUs in one local AI server without spending thousands on the platform: a Huananzhi X99-F8D Plus dual-socket board bought new for about $260, with six PCIe slots (three x16, three electrically x8), two used Xeon E5-2680 v4 CPUs, six Tesla P100 16GB cards for 96GB of VRAM, 64GB RAM, a 1600 watt supply with each card limited to 175 watts, Linux and llama.cpp. 'About $2,500 total, versus roughly $25,000 for a Blackwell workstation with the same VRAM.' Comment 'Ai' for the full rig masterclass.
The parts list is honest. Newegg's listing for the X99-F8D Plus shows three PCIe 3.0 x16 slots and three x8, and the bare board sells around $249, so six slots for about $260 holds. Six 16GB cards do total 96GB, and he is right that for running a model, as opposed to training one, slot speed matters little once the model is loaded. He also names real trade-offs and discloses an AI-generated call to action, which is more candor than most. The stretch is the comparison. A Tesla P100 is a 2016 data-center card with no fans of its own. Ninety-six gigabytes spread over six of them is not the same thing as 96GB on one current card: a model split six ways runs at the pace of old, slow chips plus the overhead of passing work between them. An RTX PRO 6000 Blackwell, the 96GB card he is alluding to, has listed between about $9,000 and $16,000 this year, so a $25,000 workstation is plausible, but it will be many times faster. 'Tokens per second will be fine' is never given a number. Add roughly a kilowatt of power draw, loud blower fans, and no warranty on used parts. We did not verify his $2,500 total or current P100 prices.
What holds up
- Caption (which carries the full script) read; contact sheet viewed.
- Board specs and price checked on Newegg listings: 3 x PCIe 3.0 x16 plus 3 x x8, bare board about $249.
- RTX PRO 6000 Blackwell 96GB pricing checked: roughly $9,000 to $16,000 across 2026 listings.
- Tesla P100 used prices and the builder's speed were not verified; no benchmark is shown.
What doesn’t
- 'Same VRAM' hides a large speed gap between 2016 and 2026 hardware.
- No tokens-per-second figure is given.
- Used, passively cooled server cards need added cooling and carry no warranty.
- 'Comment Ai' leads to a group or a Zoom call.
The catch
A real, cheap way to get a lot of GPU memory, compared against a machine that is far faster. Memory matches; performance does not.
How to actually do it
- Decide what you need to run first; most small-office uses fit a hosted model at a few dollars a month.
- If you need local, price one or two used 24GB cards before a six-card rig; simpler, quieter, faster per card.
- If you build the six-card rig, budget for a 1600 watt supply on its own circuit, cooling shrouds and the power bill.
- Ask any builder for tokens per second on the exact model you plan to run.
The build is real and the price is roughly right for the board. Buy it for cheap memory and tinkering, not for workstation-class speed.
- Confidence
- Medium
- Posted by
- a local-AI hardware creator (small account); reel sent to Buddy Sun Nov 15, 2026, 10:31 PM
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.