Corerope

02The survey

Stacked levers reach 1000x only against 2024

A roughly 1000x cut in cost and energy per solved task, at fixed quality, is credible against a 2024 baseline (a dense ~405B transformer in BF16 on H100s) but not, on current evidence, against the late-2026 frontier.

Known levers, against 2024
~700xIn dollars per solved task; ~2,000x in energy
Known levers, against late 2026
~35xIn dollars per solved task; ~80x in energy
The market alone, 2024 to now
14x to 471xPublished price declines at constant capability, compounded

Built from research/01-report-path-to-1000x.md (edit)

Chapter 1

Two baselines and one metric: dollars and joules per solved task

The "1000x" claim only means something once the baseline, the unit and the quality bar are fixed. This report measures dollars and joules per solved task at a fixed quality level on a fixed task suite, holding constant a per-user speed target (interactivity), a context-length distribution, and a power boundary that includes prefill GPUs, decode GPUs and idle power.

At near-equal scores, output per task differs 4.4x

Output tokens per task on the Artificial Analysis Intelligence Index v5 (October 2026, max effort). A cheaper token buys nothing if a model spends more of them.

Claude Opus 5.5 (max)Index score 58
~119k
Claude Fable 5.1
~78k
Claude Opus 5
~73k
GPT-6 Astra (max)Index score 57
~27k

The index's median output per task rose from ~35k (v4.3) to ~59k tokens (v5) in one revision.

Show the numbers
Output tokens per task
ModelOutput tokens per taskIndex v5 score
Claude Opus 5.5 (max)~119k58
Claude Fable 5.1~78knot given
Claude Opus 5~73knot given
GPT-6 Astra (max)~27k57

The market trend alone gives 14x to 471x from mid-2024 to November 2026

Each published rate of price decline at constant capability, compounded over the 2.4 years since mid-2024. Against 2024, much of 1000x is what the market already does.

Epoch, constant capability13x per year; 1000x in 2.7 years
a16z, LLMflation10x per year; 1000x in 3.0 years
Open-weight Pareto (Gundlach)4.6x per year; 1000x in 4.5 years
Algorithmic only, hardware removed (Gundlach)3x per year; 1000x in 6.3 years
Times cheaper at constant capability by November 2026 (log scale)

Derived: rate to the power of the years, as the survey computes it. These are prices, not costs, so margin compression inflates the headline.

Show the numbers
Market trend estimates
EstimatePer yearBy November 2026Years to 1000x
Epoch, constant capability13x471x2.7
a16z, LLMflation10x251x3.0
Open-weight Pareto (Gundlach)4.6x39x4.5
Algorithmic only, hardware removed (Gundlach)3x14x6.3

Read the full sectionSource: research/01-report-path-to-1000x.md, lines 5 to 25

Chapter 2

Every lever attacks one of four terms, and the binding term rotates

Cost per solved task factors into four terms: tokens per task, bytes and FLOPs per token (weight bytes, KV bytes and arithmetic), cost or energy per byte and per FLOP (the hardware), and utilization (batching, scheduling and idle time). The physics explains why overlap is the rule.

On H100-class silicon, reading one 8-bit weight from HBM costs about 90x the two FLOPs it feeds, and the batch size at which amortized weight reads equal arithmetic energy is about 90 for 8-bit weights and about 45 for 4-bit (derived from arXiv 2607.26571). Above that batch a GPU becomes arithmetic-bound, except for KV reads, which grow with context and are never shared across sequences.

Read the full sectionSource: research/01-report-path-to-1000x.md, lines 27 to 35

Over 90% of a 2024 token's energy is moving bytes, not doing math

One generated token of the B2024 baseline (Llama-3.1-405B-class, BF16, 16 H100s, 64K context), at H100's measured 11.68 pJ per HBM bit and 0.52 pJ per tensor-core FLOP.

One B2024 token~9.3 J
  • Weight reads (shared by the batch) 5.4 J
  • KV-cache reads (this sequence's own) 3.2 J
  • Arithmetic 0.7 J

Derived, not measured: each token reads ~92.1 GB from HBM (its batch's share of the weights plus its own 33.8 GB of KV cache, with ~13.9 sequences fitting beside the weights). Static power is not counted.

Show the numbers
Energy of one B2024 token
TermJoulesShare
Weight reads (shared by the batch)5.4558.5%
KV-cache reads (this sequence's own)3.1634%
Arithmetic0.77.52%
Total9.31100%

Batching hides weight reads only until arithmetic binds

Per-token energy to read a Dense 70B model's weights from HBM, shared across a batch, against the 0.073 J of arithmetic each token needs. Above the crossover batch, fewer bits stop helping on a GPU.

  • 16-bit weights
  • 8-bit weights
  • 4-bit weights
  • 1.58-bit weights
  • Arithmetic per token
Show the numbers
Crossover batch by weight precision
WeightsRead energy at batch 1Crossover batch
16-bit weights13.1 J180
8-bit weights6.54 J89.8
4-bit weights3.27 J44.9
1.58-bit weights1.29 J17.7

Chapter 3

What November 2026 actually measures, lever by lever

8 families of levers, each with what has actually been measured, what it costs in quality, and what it overlaps with. Each verdict is the survey's own heading; the full section and the note behind it hold every source.

01

Low-bit numerics

FP4 is solved, ternary is not

  • FP4 QAT (W4A4)3.6x weight bytes; 1.5-3x end to end

    Quality cost
    ~1% with QAT/native; PTQ worse when overtrained
    Overlaps with
    Hardware generation, MoE memory
    Evidence
    Vendor research, production adoption
  • Ternary / 2-bit1.0-1.3x over 2-bit kernels on CPU/GPU

    Quality cost
    2-bit PTQ collapses reasoning; ternary at frontier scale unproven
    Overlaps with
    Sparsity (30-52% zeros), custom silicon
    Evidence
    Mostly vendor; Intel independent
  • KV quantizationFP8 2x; NVFP4 ~2x more

    Quality cost
    0.7-2 pts (FP8); <1% (NVFP4, vendor)
    Overlaps with
    Attention compression
    Evidence
    Independent (vLLM) plus vendor
  • 2:4 weight sparsity1.1-1.3x alone

    Quality cost
    ~2-3%
    Overlaps with
    Quantization (quality adds)
    Evidence
    Vendor plus academic

02

MoE is a ~7x lever at equal quality, realized only at hyperscale batch

  • MoE at ~3% activation~7x efficiency leverage; 5.2x vs naive TP serving

    Quality cost
    None by construction; ~1.3x less knowledge per total parameter
    Overlaps with
    Batching, low-bit memory
    Evidence
    Academic law, open reproduction
  • Activation sparsity / early exit1.5-1.8x at batch 1; ~1x at scale

    Quality cost
    Minimal
    Overlaps with
    MoE, distillation
    Evidence
    Academic

03

Distillation and capability density

a fast trend against a hard knowledge floor

  • Distillation / density~10x/yr fewer parameters at fixed benchmark

    Quality cost
    Brittleness, hallucination, 2-4x tokens
    Overlaps with
    MoE, routing
    Evidence
    Independent (Epoch, AA) plus academic

04

Attention and KV

~50x at a million tokens, ~1x to 2x at 8K

  • Compressed / hybrid attention6-7x decode at 128K; ~50x KV at 1M

    Quality cost
    Multi-hop risk at scale (MiniMax)
    Overlaps with
    KV quantization; prefix caching
    Evidence
    Vendor reports, price evidence

05

Decoding

speculation buys latency, diffusion buys speed but not frontier quality

  • Speculation / MTP1.38x at batch 64; ~3x at 70-140 tok/s/user

    Quality cost
    Lossless
    Overlaps with
    Batching, low-bit, diffusion
    Evidence
    Independent plus vendor
  • Diffusion LLMs5-15x per-user speed; ~1-2x cost

    Quality cost
    10-20 pts on reasoning (same lab)
    Overlaps with
    Speculation
    Evidence
    Mostly vendor, self-reported

06

Tokens per task

the lever moving the wrong way

  • Reasoning efficiency~2x (research); 4.4x spread at frontier

    Quality cost
    ±3 pts
    Overlaps with
    Multiplies with per-token levers
    Evidence
    Independent (AA) plus academic
  • Routing / effort control~8x cost for -4 pts

    Quality cost
    Bounded by hard-task share
    Overlaps with
    Density
    Evidence
    Independent (AA)

07

Serving software

1.2x to 2.8x in six months on the same silicon

  • Serving software1.2-2.8x on same GB300 in 6 months

    Quality cost
    None
    Overlaps with
    MoE, MTP
    Evidence
    Audited (MLPerf)

08

Hardware

GPUs compound; exotic silicon mostly buys speed, not cost

  • GPU generation3.1x (H100 to B200); 2-5.4x (GB300 to Rubin) per MW

    Quality cost
    None
    Overlaps with
    Embedded FP4/LUT
    Evidence
    Independent (InferenceX)
  • SRAM / wafer-scale3-6x per-user speed; cost "comparable"

    Quality cost
    None
    Overlaps with
    Hardwired
    Evidence
    AA speeds, vendor cost
  • Hardwired weights~50-74x per-user speed; ~1.4x cost at R1 scale

    Quality cost
    Aggressive 4-bit, unmeasured
    Overlaps with
    Low-bit, MoE (30 tape-outs)
    Evidence
    Vendor plus EE Times hands-on
  • Analog in-memory~10x power (simulated)

    Quality cost
    Needs noise-aware training
    Overlaps with
    Low-bit
    Evidence
    Simulation only
  • PhotonicsNone published

    Quality cost
    Unknown
    Overlaps with
    Attention alternatives
    Evidence
    Vendor demo only

Source: research/01-report-path-to-1000x.md, lines 37 to 112

Chapter 4

The multiplier budget: a million on paper, about a thousand in practice

A useful check is the ceiling the physics allows (derived). An MoE with ~20B active parameters in FP4, batch-amortized weights and sparse KV reads does ~40 GFLOP of weight math per token, about 20x less than the baseline's 0.8 TFLOP, and at H100 coefficients its energy floor is roughly 0.03 to 0.05 J per token, about 200x below B2024's ~9 J, before any hardware improvement. Realized model-side gains land well below that ceiling because of kernel efficiency, expert-parallel communication, prefill and idle power.

Multiplied naively the levers promise a million; overlap-adjusted, about a thousand

Each lever's standalone multiplier (low to high, dot at central), grouped by the term it attacks, then the overlap-adjusted subtotals and totals. Hollow dots are the naive product of every central row.

Model side, per token
FP4 to ~3-bit LUT or ternary expertsRubin LUT is ~1.36x fewer bytes than MXFP4
MoE from ~3% to ~1.5% activationSublinear efficiency leverage
Frontier distillation plus retrieval offload of knowledgeKnowledge floor; brittleness risk
Activation sparsity / early exitNear zero at scale
KV from FP8 to FP4 / 2-bitKV already ~2% of GQA baseline, so its share is small
Hybrid linear layers on top of compressed-sparse attentionSame KV term
Model-side per token (overlap-adjusted)Naive central ~11
Utilization: serving and idle compute
Serving maturity on new architecture plus DSpark-class draftingMTP already in baseline
Hardware
GB300 to Rubin (cost)Energy central ~3x
Hardwired / SRAM tier for routed student (cost)Energy central ~2x; up to ~10x on routed share
Tokens per task
Reasoning-token efficiency (cost-aware RL, adaptive thinking)Frontier spread 4.4x is the existence proof
Routing cascade and effort controlCapped by hard-task share; overlaps the row above
Tokens-per-task subtotal (overlap-adjusted)
All levers together
Naive product of all central rows
Overlap-adjusted total, cost per solved taskHigh is improbable: every lever must land
Overlap-adjusted total, energy per solved task
Times cheaper per solved task (log scale)

Every compounding factor here is the survey's inference: no source has measured these levers together at frontier scale.

Show the numbers

Table B: against B2026 (DeepSeek-V4-Pro-class on GB300 NVL72, late-2026 software), fixed late-2026-frontier quality

Table B: against B2026 (DeepSeek-V4-Pro-class on GB300 NVL72, late-2026 software), fixed late-2026-frontier quality
LeverLowCentralHighOverlap treatment
FP4 to ~3-bit LUT or ternary experts1x1.3x2xRubin LUT is ~1.36x fewer bytes than MXFP4
MoE from ~3% to ~1.5% activation1.1x1.4x1.8xSublinear efficiency leverage
Frontier distillation plus retrieval offload of knowledge1.2x2x4xKnowledge floor; brittleness risk
Activation sparsity / early exit1x1.05x1.2xNear zero at scale
KV from FP8 to FP4 / 2-bit1.3x2x3xKV already ~2% of GQA baseline, so its share is small
Hybrid linear layers on top of compressed-sparse attention1.1x1.5x2.5xSame KV term
Model-side per token (overlap-adjusted)1.2x2x5xNaive central ~11
Serving maturity on new architecture plus DSpark-class drafting1.2x1.7x3xMTP already in baseline
GB300 to Rubin (cost)1.5x2x4xEnergy central ~3x
Hardwired / SRAM tier for routed student (cost)1x1.3x3xEnergy central ~2x; up to ~10x on routed share
Reasoning-token efficiency (cost-aware RL, adaptive thinking)1.3x2x4xFrontier spread 4.4x is the existence proof
Routing cascade and effort control1.5x3x6xCapped by hard-task share; overlaps the row above
Tokens-per-task subtotal (overlap-adjusted)2x4x10x
Naive product of all central rows420x
Overlap-adjusted total, cost per solved task4x35x1,800xHigh is improbable: every lever must land
Overlap-adjusted total, energy per solved task6x80x5,000x

Table A: against B2024 (dense ~405B, BF16, H100, 2024 software), fixed 2024-frontier quality, ~64K agentic mix

Table A: against B2024 (dense ~405B, BF16, H100, 2024 software), fixed 2024-frontier quality, ~64K agentic mix
LeverLowCentralHighOverlap treatment
FP4 weights and activations (QAT)1.5x2.5x3.6xWeight group
Sub-4-bit (LUT / ternary) beyond FP41x1.2x2xWeight group; GPU kernels cap it
MoE at ~3% activation3x6x10xWeight group; needs wide-EP batch
Capability density (data, overtraining, distillation)4x10x30xWeight group; overlaps MoE on active parameters
Activation sparsity / adaptive depth1x1.1x1.5xWeight group; erodes at batch
Weight-group subtotal (overlap-adjusted)15x60x200xNaive central ~200
Attention architecture (MLA, hybrid, compressed-sparse)3x8x20xKV group
KV quantization (BF16 to FP8/FP4)2x3x4xKV group, same term
KV-group subtotal (overlap-adjusted)4x12x40xNaive central ~24
Model-side per token (byte-share combination plus batch synergy)10x30x80xShares ~65% weights / ~35% KV at capacity-limited batch
Serving software and scheduling1.5x2.5x5xUtilization group
Speculation / MTP1.2x1.6x3xUtilization group; same idle compute as batching
Diffusion decoding1x1x2xNot credited (overlaps speculation; quality gap)
Utilization subtotal1.5x2.5x4x
GPU generation H100 to Rubin (cost)3x5x8xEnergy central ~10x; FP4 counted above
Specialized silicon for routed share (cost)1x1.2x2xEnergy central ~2x
Tokens per task (reasoning efficiency, routing)0.7x1.5x3xB2024 is a non-reasoning model, so small reasoning replacements can spend more tokens
Naive product of all central rows1,000,000x
Overlap-adjusted total, cost per solved task30x700x15,000xHigh requires every lever at its best
Overlap-adjusted total, energy per solved task60x2,000x40,000xHardware energy improves faster than $/token

Read the full section, with tables A and BSource: research/01-report-path-to-1000x.md, lines 114 to 166

Chapter 5

Honest odds, and the problems nobody has solved

Against B2026 at late-2026 frontier quality, the central estimate is ~35x in dollars and ~80x in energy, with a plausible band of ~10x to ~100x. Against whatever the frontier is when the program finishes, the advantage over a typical deployment stays roughly 10x to 50x, about the size of the cost-per-task spread already visible among top models. Reaching 1000x against the contemporaneous frontier would need a hardware discontinuity, and the only candidate, hardwired silicon at frontier scale, shows ~1.4x in cost by its own vendor's figures, not 10x.

Whether "1000x cheaper" is achievable depends mostly on which year's frontier serves as the baseline. Against 2024 it is an accounting exercise the market is already completing; against late 2026 it is out of reach for any single program on current evidence. The realistic goal is a sustained 10x to 50x lead over typical deployment at frontier quality, worth one to two years on the price curve.

Read the whole surveySource: research/01-report-path-to-1000x.md, lines 194 to 200