Skip to content
MINGYANG FUNDIndependent investment research
AI / INVESTMENT INTELLIGENCE

Read the paper. Follow the economics.

Plain-language research briefs connecting technical mechanisms to workloads, infrastructure and value-chain exposure.

10 paper briefs · Click the paper card to open the original. Investment implications below are analytical hypotheses, not claims of booked revenue.

01Custom silicon
2026-08-25

Maia 200: make data movement programmable

Microsoft's accelerator paper starts with a practical problem: a chip can perform many calculations yet still sit idle waiting for data. Maia explicitly programs how data moves through specialized memories and processing engines. The aim is to keep inference hardware productive while reducing the cost of moving information.

Read original paper ↗
INVESTMENT TRANSMISSION
Custom silicon / foundry Potential share shift

Successful in-house inference chips can retain more infrastructure economics inside a cloud platform and change the mix of purchased GPUs.

MSFT · TSM · NVDA
Memory / networking System bottleneck

More arithmetic requires effective memory and interconnect throughput; benchmark peak FLOPS alone misses these constraints.

MU · AVGO
Findings, limits & what to watch

Original paper: Maia 200: A Software Defined Dataflow System for Large-scale AI Acceleration

  • Introduces a software-defined dataflow architecture that coordinates memory and computation explicitly.
  • Describes low-precision inference hardware with high-bandwidth memory and large-scale parallelism.

Where the result may not travel

Architecture specifications and vendor-reported efficiency do not establish broad customer adoption or matched-workload total cost against competing platforms.

Evidence to watch

Compare accepted tasks per dollar, utilization and porting effort on the same model. Track deployment scale rather than peak FLOPS alone.

Sherry Xu, Marco Heddes, Jackson Peng et al.

02Custom silicon
2026-06-14

Google TPU: five generations of scale

Google reviews five TPU generations to explain how a relatively stable chip architecture can scale into much larger training systems. The paper treats memory, communication, fault recovery and power efficiency as part of the product, rather than judging an accelerator only by its calculation speed.

Read original paper ↗
INVESTMENT TRANSMISSION
Cloud / custom silicon Potential integration advantage

Control of chips, software and the data center can improve utilization and reduce dependence on a single external platform.

GOOGL · NVDA
Memory / optical networking Higher system importance

As cluster scale rises, data movement and resilience can limit delivered compute even when chips are available.

MU · GLW · AVGO
Findings, limits & what to watch

Original paper: Google's Training Supercomputers from TPU v2 to Ironwood: Architectural Stability, Scale, Resilience, Power Efficiency, and Sustainability Across Five Generations

  • Documents improvements in memory capacity, bandwidth and system scale across successive TPU generations.
  • Explains how optical circuit switching and fault-recovery mechanisms support useful work at large scale.

Where the result may not travel

This is a Google-authored history of its own systems. Cross-generation improvements do not prove superior customer economics against every current GPU workload.

Evidence to watch

Follow external TPU customer adoption and measured training time at a fixed model quality. Supplier names here are industry exposures, not a disclosed Ironwood bill of materials.

Norman P. Jouppi, Sridhar Lakshmanamurthy, Cliff Young and David Patterson

03Demand & adoption
2026-01-15

OpenRouter: where the tokens actually go

This study uses real interactions routed through OpenRouter to examine AI usage across models, tasks, places and time. It shifts the question from which model wins a test to where developers and users choose to send work. That is useful for testing whether better capability becomes recurring demand.

Read original paper ↗
INVESTMENT TRANSMISSION
Inference / cloud Demand validation

Persistent paid workload growth could offset falling token prices; volume alone cannot establish this offset.

AMZN · MSFT · GOOGL · NVDA
Applications / distribution Bargaining-power test

Routing across models can weaken a single model vendor's hold on customers while increasing the value of workflow ownership.

MSFT · GOOGL
Findings, limits & what to watch

Original paper: State of AI: An Empirical 100 Trillion Token Study with OpenRouter

  • Analyzes more than 100 trillion tokens of interactions on one multi-model inference platform.
  • Studies usage patterns during the move toward reasoning models and multi-step inference.

Where the result may not travel

OpenRouter users are a selected population. Token volume is not revenue, profit, active users or the entire AI market; free traffic and differing tokenizers complicate comparisons.

Evidence to watch

Track comparable routed token volume alongside price, paid retention and workload mix. Do not turn developer nationality into user geography.

Malika Aubakirova, Alex Atallah, Chris Clark, Justin Summerville and Anjney Midha

04Agents & reliability
2025-03-18

METR: how long can an agent work?

METR measures task difficulty in a familiar unit: the time a skilled person typically needs. Its 50% completion horizon describes tasks an AI can finish half the time. This makes progress easier to interpret, while exposing the gap between a convincing short demo and reliable autonomous work.

Read original paper ↗
INVESTMENT TRANSMISSION
Software / services Conditional automation

Longer reliable workflows could change seat demand and labor intensity, but supervision and verification remain economic costs.

MSFT · CRM · ACN
Inference / cloud Potential workload growth

Autonomous tasks can require more tool calls, retries and tokens even if each individual inference becomes cheaper.

AMZN · GOOGL · NVDA
Findings, limits & what to watch

Original paper: Measuring AI Ability to Complete Long Software Tasks

  • Combines human task timings with model success rates across software and reasoning tasks.
  • Documents historical growth in the measured horizon and discusses limits to extrapolating that trend.

Where the result may not travel

A 50% success threshold is not production reliability. The task sample, scaffolding and human baselines constrain generalization to jobs and company-wide productivity.

Evidence to watch

Monitor high-reliability success thresholds, human review time and cost per accepted task, not just the longest demonstrated task.

Thomas Kwa, Ben West, Joel Becker et al. / METR

05Reasoning
2025-01-22

DeepSeek-R1: learning to reason

DeepSeek shows how reinforcement learning can encourage a model to check its work and change strategy on problems with verifiable answers. It also transfers reasoning behavior into smaller models. The investment question is whether extra computation at inference produces enough useful work to justify its cost.

Read original paper ↗
INVESTMENT TRANSMISSION
Inference compute More work per request

Reasoning may increase tokens and server time per query; smaller models and better accuracy can offset the cost.

NVDA · AMZN · MSFT
Applications Lower entry barrier

Accessible reasoning models may reduce model-level differentiation and shift value toward distribution, proprietary data and reliable workflows.

MSFT · GOOGL · CRM
Findings, limits & what to watch

Original paper: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

  • Reports emergent reasoning behavior from reinforcement learning without relying on human-written reasoning traces at every step.
  • Shows that large-model reasoning can help train smaller models through distillation.

Where the result may not travel

Strong math and coding results do not establish reliability in open-ended business tasks. Smaller distilled models can lose capability, and longer reasoning can increase latency and token costs.

Evidence to watch

Measure success-adjusted cost and latency under a fixed task budget. More tokens are valuable only when they improve accepted outcomes.

DeepSeek-AI, Daya Guo, Dejian Yang et al.

06Model efficiency
2024-12-27

DeepSeek-V3: doing more with less compute

V3 uses a mixture of experts so each token activates only part of a large model. It combines that approach with compressed attention memory and low-precision training. The central idea is to make each unit of useful computation cheaper through coordinated model and systems design.

Read original paper ↗
INVESTMENT TRANSMISSION
Compute / cloud Lower unit cost; demand uncertain

Cheaper useful inference can pressure price per token while making more applications economic. Total accelerator demand depends on the demand response.

NVDA · AMZN · GOOGL
Memory / networking Different bottleneck mix

Expert routing and large model weights can preserve memory and communication constraints despite lower active arithmetic.

MU · AVGO · ANET
Findings, limits & what to watch

Original paper: DeepSeek-V3 Technical Report

  • Reports 671 billion total parameters with 37 billion activated per token.
  • Combines Multi-head Latent Attention, expert routing and FP8 training to improve computational efficiency.

Where the result may not travel

Reported training compute is not a complete corporate R&D budget. Sparse activation does not eliminate the need to store model weights, move data or coordinate expert servers.

Evidence to watch

Compare utilization, interconnect traffic and total cost at matched output quality; avoid interpreting a single training bill as an industry spending forecast.

DeepSeek-AI, Aixin Liu, Bei Feng et al.

07Reasoning
2024-08-06

Test-time compute: when thinking longer pays

Instead of using the same inference effort on every question, this paper asks when a model should spend more time revising or searching for an answer. Its key point is that the best use of extra compute depends on the difficulty of the problem and the method used to check candidate answers.

Read original paper ↗
INVESTMENT TRANSMISSION
Inference / cloud Variable resource intensity

Hard tasks can become more compute-intensive while easy tasks are routed cheaply, changing capacity planning and pricing.

NVDA · MSFT · AMZN
Software / orchestration Potential value shift

Routing and verification software can capture value if it lowers cost per accepted answer without degrading reliability.

MSFT · GOOGL
Findings, limits & what to watch

Original paper: Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

  • Compares approaches for allocating additional computation when answering a prompt.
  • Finds that adapting the strategy to prompt difficulty can use the compute budget more effectively than a uniform sampling baseline.

Where the result may not travel

The result is task- and verifier-dependent. Extra search cannot always compensate for missing knowledge or weak evaluation, and end-user latency may limit adoption.

Evidence to watch

Track quality at a fixed total budget, including verifier and retry costs, and distinguish latency-sensitive products from offline work.

Charlie Snell, Jaehoon Lee, Kelvin Xu and Aviral Kumar

08Serving efficiency
2024-07-11

FlashAttention-3: keep the GPU working

The attention step in a language model moves large amounts of information between memory and calculation units. FlashAttention-3 overlaps those operations and uses Hopper hardware features more effectively. It demonstrates how software can extract more useful performance from an already-installed accelerator.

Read original paper ↗
INVESTMENT TRANSMISSION
GPU / cloud More output per installed chip

Better utilization can reduce required GPUs for a fixed workload or enable more demand at a lower price; the net hardware effect is conditional.

NVDA · AMZN · MSFT · GOOGL
Software tooling Implementation advantage

A strong compiler and kernel ecosystem can extend the economic usefulness of an accelerator platform.

NVDA · AMD
Findings, limits & what to watch

Original paper: FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

  • Uses asynchronous execution and low-precision techniques to improve attention-kernel utilization on H100 GPUs.
  • Reports gains against earlier attention implementations under the paper's hardware and workload settings.

Where the result may not travel

An attention-kernel speedup is not the same as the speedup of a complete application. Benefits vary with sequence length, batch size, numerical format and other bottlenecks.

Evidence to watch

Benchmark complete inference and training jobs, measuring tokens per second, latency and power at matched quality rather than copying kernel headline gains.

Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani and Tri Dao

09Model efficiency
2023-12-01

Mamba: a different way to remember

Mamba explores an alternative to storing and attending over every earlier token. It maintains a compact state that selectively updates as new information arrives. The architecture targets long sequences with more favorable computational scaling while trying to preserve the ability to use relevant content.

Read original paper ↗
INVESTMENT TRANSMISSION
Memory / compute Possible intensity change

If adopted at comparable quality, recurrent or hybrid designs can change the balance between cache memory, arithmetic and bandwidth per token.

MU · NVDA · AMD
Cloud / applications New feasible workloads

Cheaper long-sequence processing could expand document, audio and other streaming applications; commercial demand still needs validation.

AMZN · GOOGL · MSFT
Findings, limits & what to watch

Original paper: Mamba: Linear-Time Sequence Modeling with Selective State Spaces

  • Introduces input-dependent selection within a state-space sequence model.
  • Pairs the architecture with a hardware-aware implementation to process long sequences efficiently.

Where the result may not travel

Efficient scaling does not guarantee equal retrieval or reasoning quality on every task. Results for the studied models do not establish wholesale replacement of Transformers.

Evidence to watch

Compare long-context retrieval accuracy and total serving cost with current Transformer and hybrid systems on the same workload.

Albert Gu and Tri Dao

10Serving efficiency
2023-09-12

PagedAttention: fit more work into memory

When many people use a language model together, each request needs a growing memory of earlier tokens. PagedAttention manages that memory in smaller blocks, similar to virtual memory in an operating system. The vLLM serving system uses this approach to fit more useful work into the same hardware.

Read original paper ↗
INVESTMENT TRANSMISSION
Inference / cloud Lower serving cost

More concurrent requests per device can improve provider margins or be passed through as lower API prices.

AMZN · MSFT · GOOGL
Memory / accelerators Demand response uncertain

A fixed workload can need fewer devices, while cheaper access can expand usage. Hardware spending follows the combined effect.

MU · NVDA · AMD
Findings, limits & what to watch

Original paper: Efficient Memory Management for Large Language Model Serving with PagedAttention

  • Reduces wasted key-value cache space and supports sharing cache blocks where appropriate.
  • Demonstrates higher serving throughput than the comparison systems at comparable latency in the paper's experiments.

Where the result may not travel

The paper's 2023 baselines are historical. Actual gains depend on workload, scheduling, context length and hardware; eliminating fragmentation does not eliminate memory demand.

Evidence to watch

Measure sustained throughput under a latency service-level objective, including realistic request lengths, and track whether savings become margin or price cuts.

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang et al.