$HBM is live on Pons v2, Robinhood Chain.$HBM is live on Pons v2. Buy $HBM

HBM
WebGPUEnterprise

AI is only
as fast as
its memory.

One token from a 70B model at 4-bit reads 35 GB of weights. Bandwidth, not arithmetic, sets the speed.

From laptop to data center

  1. Thin-and-light laptopLPDDR5X-7500, 128-bit120 GB/s
  2. Apple M4 ProLPDDR5X unified273 GB/s
  3. Apple M4 MaxLPDDR5X unified546 GB/s
  4. GeForce RTX 4090GDDR6X, 384-bit1,008 GB/s
  5. GeForce RTX 5090GDDR7, 512-bit1,792 GB/s
  6. NVIDIA H100 SXMHBM3, 5 stacks3,350 GB/s
  7. NVIDIA H200HBM3E, 6 stacks4,800 GB/s
  8. AMD Instinct MI300XHBM3, 8 stacks5,300 GB/s
  9. NVIDIA B200HBM3E, 8 stacks8 TB/s
Published peak, log scale. Run the test to add yours.

Inside an HBM package

Cross-section of an AI accelerator package: two HBM stacks of twelve DRAM dies on base logic dies sit beside the compute die on a silicon interposer, above the package substrate.GPU / accelerator dieDRAM dies, 12–16 highTSVs through every dieBase logic dieMicrobumpsSilicon interposer1,024 to 2,048 wires per stackPackage substrate
HBM generations
GenerationInterfacePin speedPer stackStack heightCapacity
HBM2E1024-bit3.6 Gb/s460 GB/s4–8 highup to 16 GB
HBM31024-bit6.4 Gb/s819 GB/s8–12 highup to 24 GB
HBM3E1024-bit9.2–9.8 Gb/s≈1.2 TB/s8–12 high24–36 GB
HBM42048-bit8 Gb/s (JEDEC)2 TB/sup to 16 highup to 64 GB

The supply chain, by role

Supported RWA active Robinhood Stock Token. Industry company listed for its role.

AI memory

Make the DRAM dies and stack them into HBM.

  • SK hynixSupported RWA

    Largest HBM supplier. First to mass-produce 12-high HBM3E.

    Products
    HBM3E 12-high, HBM4
    Listing
    KRX 000660 · US ADS
    Price
    —
    Stock Token
    SKHY on Robinhood Chain
  • MicronSupported RWA

    US memory maker. 12-high 36 GB HBM3E, GDDR7 for graphics.

    Products
    HBM3E 12-high, GDDR7, LPDDR5X
    Listing
    NASDAQ MU
    Price
    —
    Stock Token
    MU on Robinhood Chain
  • Samsung ElectronicsIndustry company

    One of the three major DRAM makers. HBM3E and HBM4.

    Products
    HBM3E, HBM4, LPDDR5X, GDDR7
    Listing
    KRX 005930

Packaging

Put HBM stacks beside the compute die on one silicon interposer.

  • TSMCSupported RWA

    Builds most AI accelerators. CoWoS places HBM beside the compute die.

    Products
    CoWoS-S / -L, N4 / N3 logic
    Listing
    NYSE TSM (ADR) · TWSE 2330
    Price
    —
    Stock Token
    TSM on Robinhood Chain

Accelerators

Design the processors that HBM feeds.

  • NVIDIASupported RWA

    H200: 141 GB HBM3E at 4.8 TB/s. B200: 8 TB/s.

    Products
    H200, B200, B300
    Listing
    NASDAQ NVDA
    Price
    —
    Stock Token
    NVDA on Robinhood Chain
  • AMDSupported RWA

    MI300X: 192 GB HBM3 at 5.3 TB/s. MI355X: 288 GB at 8 TB/s.

    Products
    Instinct MI300X, MI325X, MI355X
    Listing
    NASDAQ AMD
    Price
    —
    Stock Token
    AMD on Robinhood Chain

Checked against Robinhood's token registry, live 2026-09-27. Stock Tokens are issued by Robinhood Assets (Jersey) Limited and are not offered to U.S. persons.

What your number means

The fastest a model can decode on your memory. Measure your machine

Model size
Weight precision
25.0tokens per second, at most8B parameters at 4-bit is 4.0 GB of weights, streamed once per token.

Educational only. Not investment advice or an offer of any security or token. Product figures are vendor specifications.