the dream is dead

Overview

Mutahar reflects on how computer hardware—especially GPUs and memory—has become significantly more expensive due to two major waves:

• Crypto (earlier GPU shortages)
• The AI boom (current and ongoing pricing pressure)

He uses his own build history as context, then explains why he’s considering upgrading again.

---

Personal build history / setup changes

• Years ago, he built a “dream computer” with two flagship GPUs in a single system:
  • An RTX card (likely RTX 3080)
  • An AMD Radeon RX 6900 XT

Why two GPUs?

The main reason was virtual machine (VM) workflows:

• He ran Linux in a VM while gaming in virtualized environments.
• His streaming setup involved passing a GPU to a VM (Linux with its own dedicated GPU).

Why he moved away from the dual-GPU approach

He no longer needed heavy VM usage, so he simplified:

• Now he’s on a single RTX 4090.

---

Why he wants a new GPU (even though he doesn’t need it for gaming)

He’s interested in the RTX 5090, but not primarily for extra gaming FPS.

His motivations are more about AI experimentation and setting up a home lab / self-hosting:

• Hosting services such as:
  • self-run search engines
  • photo backups
  • file hosting
  • emulation frontends
• Running local downloadable AI models at home

AI and emulation progress

He points to progress in emulation aided by AI—citing examples like the PS5 emulator reaching “Red Dead Redemption 1” at roughly ~12 FPS—and suggests AI models can help with parts of emulation automation (the “boring stuff”).

---

Model costs driving hardware desire

He argues that running newer, larger AI models is increasingly expensive.

• He references model families such as GLM 5.3, claiming they’re too costly without extremely expensive hardware (on the order of hundreds of thousands up to ~a million dollars).
• He wants to use models such as DeepSeek.
• He estimates that the relevant hardware costs could be around $12,000–$15,000 (noting the estimate/timing he discusses may have shifted).

Frustration about price trends

He expresses frustration about paying premiums for hardware while expecting prices to fall eventually—yet seeing prices continue rising.

---

Evidence of RTX 5090 price inflation (forums / price trackers)

To support the claim that prices are climbing, he references:

• NeoGAF
• PCPartPicker
• other community price tracking

Example timeline he provides

• Launch price: about $2,000 (he corrects himself mid-sentence)
• By June 2026: up to about $4,200
• By August 2026: up to about $4,700

He frames this as over 130% above launch (i.e., more than ~2×).

---

The “AI hardware tax” argument

He argues Nvidia is unlikely to reduce GPU pricing back down once consumers are effectively “forced” to buy for AI workloads.

• He believes even low-end GPUs have become too expensive.
• He suggests prices may stay high as a market floor.

Best-case path out (as he speculates)

He describes a potential escape route:

• Cheaper open/alternative models (examples: DeepSeek, GLM, etc.) become accessible enough that consumers shift away from only the most expensive closed systems.
• If AI models become cheaper to run, demand for top-tier GPUs may slow.

---

Future direction: thin clients / data centers (and why he dislikes cloud AI)

He predicts a shift where major providers (hyperscalers / data centers) may offer thin-client boxes:

• The local device connects to remote compute via web tools / browsers.

Cloud AI critique

He specifically criticizes cloud AI and voices strong dislike for relying on services such as:

• Gemini
• ChatGPT
• Claude

His preference is self-hosting—running models locally rather than relying on internet-based inference.

---

Practical tooling example: LM Studio + local LLMs

He mentions LM Studio as a way to download and run local models.

He lists examples of models he can download in LM Studio, including:

• Qwen 3.6
• Qwen
• Muse
• Laguna
• Gemma (as referenced in the subtitles)

Example uses he describes

• Asking a local model for tasks like price history of RTX 50 series and tracking increases “since launch.”
• Running smaller models (he mentions around ~9B parameters as an example) on capable consumer hardware.
• Using local AI for research workflows, such as finding obscure laws and returning links—saving time compared to traditional search.

---

Key conclusion

He summarizes the situation as:

• Technology prices have escalated beyond normal reason.
• Industry sentiment suggests things could get worse into the late 2020s.

He ends by saying he mainly wanted to vent.

---

Main speakers / sources

• Speaker: Mutahar (the creator “Mutahar / it’s me Mutahar”)
• Referenced sources/tools: NeoGAF, PCPartPicker, LM Studio
• Mentioned companies/models/services: Nvidia, Anthropic, OpenAI, Gemini, ChatGPT, Claude
• AI model references: DeepSeek, GLM 5.3, Qwen, Gemma