Contact

Insights

Research on GPU pricing, compute markets, GPU financing and LLM agents, with every figure sourced and dated.

Follow Rillor on X

Latest

RSS feed

All research

Filter by pillar or type, search by topic or period, and sort by date.

  1. ExplainerDataset licenses compared: open, research-only and commercial termsDataset licensing and usage rights compared from the license texts: what open, noncommercial, research-only and commercial terms say about purpose, shared changes and trained models.Data marketplace7 min readPDF report
  2. Methods noteEU AI Act obligations as a register: providers and deployersProvider and deployer duties in Regulation (EU) 2024/1689, read into register rows by role, system scope and date, with each change from the Digital Omnibus traced.Document and regulation intelligence7 min readPDF report
  3. Methods noteEvaluation sets for agents: building questions with known answersAn agent training dataset and an agent evaluation set share one record: task, starting state, tools and a checkable known answer. How to write, verify and keep them apart.Data services7 min readPDF report
  4. Methods noteFrom GPU-hour price to cost per training tokenHow to calculate LLM training cost from a GPU-hour price: compute per token, dense peak FLOPS, utilization in published runs and a worked example with dated public prices.GPU and neocloud pricing7 min readPDF report
  5. Research reportH100 rental price history, 2023 to 2026An H100 price history from dated public sources: what list prices, rental indexes and surveyed contracts show for NVIDIA H100 GPUs from 2023 to November 2026.Rillor Compute Index7 min readPDF report
  6. Methods noteHow to compare GPU system prices across configurationsTwo GPU servers rarely match part for part. Comparing GPU server prices across configurations takes matched builds, line-item adjustment, price per unit of performance or hedonic regression.GPU systems pricing7 min readPDF report
  7. Research reportLanguage models as annotators: agreement with people in published studiesWhat published studies find when an LLM is used as an annotator: agreement with crowd workers and experts, task by task, and how to check model labels.Labeling and annotation7 min readPDF report
  8. ExplainerPermission-aware retrieval for internal assistantsRAG access control for internal assistants: how retrieval returns only what the asking user may read, from per-document filters to delegated tokens, and what published leaks show.Agentic systems7 min readPDF report
  9. Research reportPoisoned memory and retrieval: what published attacks showRAG poisoning and memory poisoning plant text that superintelligence (SI) agents later retrieve and act on. What published attacks, disclosed cases and tested defenses show.Agentic security7 min readPDF report
  10. ExplainerPreparing data for language model training: a buyer's map from source to deliveryData preparation for LLM training, stage by stage: extraction, filters, deduplication, personal data, decontamination, mixing and what a buyer should receive.Data services7 min readPDF report
  11. Research reportQuality filtering for language model training data: heuristics and classifiersHow pretraining data filtering works, from heuristic rules and language identification to perplexity and quality classifiers, what dataset ablations measured and what each filter takes out.Data services7 min readPDF report
  12. Methods noteTesting when an agent should hand off to a personHow to test agent escalation to a human as one decision with two errors, using paired cases, a confusion table, checks on the handoff itself and repeated runs.SI evaluation7 min readPDF report
  13. Research reportWhat audits found about USAspending and FPDS data qualityWhat GAO and inspector general audits from 2010 to 2026 found about USAspending data quality: dollar errors, miscoded fields, nonreporting agencies, unreported other transaction agreements and duplicate subawards.Public procurement records7 min readPDF report
  14. ExplainerCeph block, object and file storage: which workload uses whichCephFS vs RBD vs RGW, from Ceph's Umbrella 21.2.0 documentation: which workloads need a block device, a shared file system or S3-compatible object storage, and why.Distributed storage7 min readPDF report
  15. ExplainerHow agent-based market simulation worksAgent-based market simulation builds a market from simulated traders and its rules, then checks the result against real data or a known answer. LLM agents now join as traders.Market simulation7 min readPDF report
  16. Methods noteHow lenders value GPU collateralHow lenders value GPU servers as collateral, from market value to orderly and forced liquidation value, and how filed GPU facilities size loans on depreciated cost and customer credit.GPU financing7 min readPDF report
  17. Research report2023 to 2026How LLM forecasts compare with superforecasters and marketsWhat LLM forecasting benchmarks show to November 2026: top-ranked entries statistically indistinguishable from superforecasters on ForecastBench, and models mostly level with or behind professional forecasters and market prices.Forecasting and market intelligence7 min readPDF report
  18. Research reportJanuary 2025 to November 2026How memory prices reach the cost of a GPU serverDated listings, filings and contract price forecasts show how memory costs reach a GPU server's price, and what lies behind the reported NVIDIA server price increase.GPU systems pricing7 min readPDF report
  19. Methods noteHow to compare GPU generations on price per FLOPGPU price performance over time depends on the precision counted, peak or delivered throughput and the price used. This note compares data-center GPU generations on one stated basis.Rillor Compute Index7 min readPDF report
  20. ExplainerLevel 1, 2 and 3 market data explainedLevel 2 vs level 3 market data: the size at each price, or every order. Venue specifications show what each level carries and which studies need it.Market data7 min readPDF report
  21. ExplainerMIG, MPS or time-slicing: sharing GPUs on KubernetesMIG vs time-slicing vs MPS on Kubernetes: what each GPU sharing method isolates, what measurements show, how the device plugin exposes it and how to size a slice.Kubernetes platforms7 min readPDF report
  22. Research reportPaid early access to market-moving informationWho gets market-moving news first and what a head start is worth: enforcement actions, agency decisions and studies, from the Michigan survey's two-second tier to 2026's Truth API.Public communications and markets7 min readPDF report
  23. ExplainerPerformance bonds, standby letters of credit and escrow comparedPerformance bond vs standby letter of credit for an equipment purchase, with the buyer's escrowed deposit beside them: what triggers each payment and what can stop it.Forward agreements7 min readPDF report
  24. ExplainerSecurity risks in self-hosted LLM deploymentsSelf-hosted LLM security in six documented failure modes: exposed inference servers, serving software flaws, model files that run code, shared KV caches, poisoned documents and GPU container escapes.Private SI cloud7 min readPDF report
  25. Research reportWhat causes data-center outages: surveys and incident reviewsUptime's 2025 survey puts power behind 45% of impactful outages. Published incident reviews show how faults in power, cooling and water spread, and what set recovery time.Data-center capacity7 min readPDF report
  26. ExplainerWhat to do before selling surplus GPU serversBefore you sell used GPU servers: save the health logs, find every data store, sanitize and certify each drive under NIST SP 800-88 Rev. 2 and screen the buyer.GPU systems7 min readPDF report
  27. ExplainerAgentic RAG and how it differs from standard RAGAgentic RAG vs RAG: standard RAG retrieves once and answers, while agentic RAG lets the model plan searches, grade what it found and search again, at a higher cost.Agentic systems7 min readPDF report
  28. Research reportCopyright and model training: what U.S. courts and the Copyright Office have saidTraining data copyright in U.S. law, from the statute and the Copyright Office's training report to court rulings on headnotes, books, news, images and code, each with its date.Acquisition and licensing7 min readPDF report
  29. ExplainerData labeling vs data annotation: label types for text and documentsHow standards and tools use the two words, and the label types for text and documents: classes, entities and spans, relationships, ratings, page regions and tables, with their storage formats.Labeling and annotation7 min readPDF report
  30. ExplainerDe-identification, pseudonymization and anonymization comparedDe-identification vs anonymization vs pseudonymization: what HIPAA, the GDPR and NIST each mean by the terms, how EU law tests anonymity and why removing names is not enough.Data services7 min readPDF report
  31. ExplainerFine-tuning data formats: instruction, chat and preference recordsData preparation for LLM fine-tuning starts with the record: instruction, chat, preference or unpaired feedback records, how each is laid out, which tokens are trained and how files are checked.Data services7 min readPDF report
  32. ExplainerHow data marketplaces for model training workHow a training data marketplace works: listings, samples, prices, licenses and delivery on the large cloud platforms, and the rules for data intermediaries and data brokers.Data marketplace7 min readPDF report
  33. ExplainerPrompt injection vs jailbreak: two different failuresA jailbreak defeats a model's safety training. A prompt injection makes an application obey untrusted text. How to tell them apart, and why it matters for superintelligence (SI) agents.Agentic security7 min readPDF report
  34. ExplainerQuestions to ask before you deploy a vendor's agentQuestions to ask an agent vendor about testing: model versions, test sets, repeated runs, escalation, prompt injection, logs and who tested, with a public source behind each.SI evaluation7 min readPDF report
  35. ExplainerRobots.txt and model training crawlers: what an opt-out coversWhat a robots.txt opt-out from model training reaches: which crawler, which use, which time and which copy, from the protocol, crawler documentation, compliance studies and the law.Data services7 min readPDF report
  36. Methods noteTracking SAM.gov notices, amendments and attachmentsThe SAM.gov Opportunities API returns only the latest active version of a notice. Keeping amendments and attachments takes version capture, document text extraction and OCR checks.Public procurement records7 min readPDF report
  37. Research reportWhat court records show about fabricated citations in LLM outputCourt opinions and sanctions orders from 2023 to 2026 show what judges found in hallucinated legal citations and what they ordered in response.Document and regulation intelligence7 min readPDF report
  38. ExplainerWhat data is used to train language modelsWhat data is used to train LLMs, stage by stage: filtered web crawls, code, papers and books, then demonstrations, preference pairs, synthetic and checkable tasks, and what must be disclosed.Data services7 min readPDF report
  39. ExplainerWhy cloud GPU launches fail and how reservations hold capacityInsufficientInstanceCapacity and other GPU launch errors come from two separate checks, quota and capacity. Provider documents show what each error means and on what terms reservations hold GPUs.GPU and neocloud pricing7 min readPDF report
  40. ExplainerAlgorithmic pricing, collusion and the law in 2026What U.S. enforcers, two appeals courts, the RealPage decrees, four state statutes and European and UK authorities say about algorithmic pricing and collusion, and what agent experiments add.Market simulation7 min readPDF report
  41. ExplainerColocation pricing per kW: how the monthly bill is builtWholesale colocation is quoted per kW per month. Lease structure, power pass-through, escalators, one-time fees and the billing start date set the bill, as public filings show.Data-center capacity7 min readPDF report
  42. Research reportDeep learning on limit order books: what benchmarks showBenchmark studies of deep learning on limit order books find lower scores on new stocks and periods than on FI-2010, and predictability that depends on a stock's tick size.Market data7 min readPDF report
  43. Research reportGPU depreciation and useful life in filings and market evidenceFilings put GPU and server useful lives at up to six years. This report compares those book lives with tax rules, failure studies, rents for older GPUs and resale evidence.Rillor Compute Index7 min readPDF report
  44. Methods noteLease or buy GPU servers: how to compare the costHow to compare a GPU lease to own, a fair-market-value lease and a loan-financed purchase on after-tax cost, and how 2026 tax rules, ASC 842 and residual value move it.GPU financing7 min readPDF report
  45. Methods noteLook-ahead bias in LLM backtests and how to test for itA language model can already know the outcomes in a historical backtest. Published tests detect this look-ahead bias, and point-in-time models and dated inputs reduce it.Forecasting and market intelligence7 min readPDF report
  46. ExplainerPrivate LLM deployment: who controls whatOn-premises LLM deployment, colocation, dedicated GPU capacity and managed APIs each place the hardware, facility, data, model weights, logs and operations with different parties.Private SI cloud7 min readPDF report
  47. Research reportSelf-hosted S3-compatible storage: Ceph RGW and the alternativesMinIO alternatives for self-hosted S3-compatible storage, compared from each project's own documentation and license as of 8 November 2026: Ceph RGW, SeaweedFS, RustFS, Garage, Apache Ozone and OpenStack Swift.Distributed storage7 min readPDF report
  48. ExplainerSlurm or Kubernetes for GPU clusters: how to chooseSlurm vs Kubernetes for GPU clusters: the unit each scheduler places, how a training job starts all at once, where each fits and how one cluster runs both.Kubernetes platforms7 min readPDF report
  49. ExplainerWhat buyers can do when GPU systems arrive lateRemedies for late delivery of goods under the Uniform Commercial Code and the international sales convention, applied to GPU systems delivered in tranches, with terms from public filings.Forward agreements7 min readPDF report
  50. ExplainerWhat posted GPU cloud prices tell you, and what they do notA posted GPU price is a listing for one configuration, term and moment. Comparing two of them takes seven adjustments, and none turns a listing into a transaction.GPU and neocloud pricing5 min readPDF report
  51. ExplainerWhat sets the price of a used H100 serverA used H100 server price depends on what the figure covers and who states it, then on condition, warranty, diagnostic records, software licenses and export screening.GPU systems pricing7 min readPDF report
  52. Research reportWhat studies find on social media posts and market movesWhat published studies find on social media posts and the stock market: effects that depend on topic, tone, asset and data frequency, null results, and moves that fade within days.Public communications and markets7 min readPDF report
  53. ExplainerWhat to check before buying a used GPU serverOn Ampere and later GPUs, nvidia-smi cannot clear lifetime ECC counts. What to read and test on a used GPU server, and the warranty, drive and export checks.GPU systems7 min readPDF report
  54. ExplainerDue diligence questions to ask a data vendorData vendor due diligence, question by question: who the vendor is, where its data comes from, how it checks its suppliers and how to test its answers.Data marketplace7 min readPDF report
  55. ExplainerExtracting data from SEC filings: XBRL and textWhich parts of a 10-K, 10-Q or 8-K must carry XBRL tags, what the data.sec.gov APIs return, and which sections stay as text to extract.Document and regulation intelligence7 min readPDF report
  56. ExplainerGPU cloud costs beyond the hourly rateThe hidden costs of GPU cloud from dated price pages: data transfer, storage, IP addresses and support, and the switching charges the EU Data Act ends from 12 January 2027.GPU and neocloud pricing7 min readPDF report
  57. Research reportHow reliable are LLM agents on repeated runsBenchmarks from 2024 to 2026 show how often LLM agents succeed on every repeat of one task, and how far a single run moves a score.Agentic systems7 min readPDF report
  58. Methods noteHow to count federal contract dollars without double countingBase and all options value vs obligated amount on federal contract records: which dollar field to sum, at which level, and how modifications, orders and outlays change a total.Public procurement records7 min readPDF report
  59. Methods noteIndirect prompt injection: testing every channel an agent readsIndirect prompt injection reaches superintelligence (SI) agents through what they read. A test plan for each channel: documents, web pages, email, tool output, retrieved data and images.Agentic security7 min readPDF report
  60. Methods noteNear-duplicate detection for training data: MinHash, SimHash and embeddingsHow near duplicate detection works on training data: shingles and Jaccard similarity, MinHash with banding, SimHash, embeddings, the settings that dataset papers report and how to test a threshold.Data services7 min readPDF report
  61. Research reportTraining data licensing deals: what public filings and reports discloseMost data licensing deals for model training are announced without a price. What SEC filings, results statements and reporting show about parties, scope, structure, term and revenue.Acquisition and licensing7 min readPDF report
  62. Methods noteValidating an LLM judge against human labelsHow to check an LLM-as-a-judge against human labels: chance-corrected agreement, probes for position, length and self-preference bias, re-validation when the judge changes, and corrected final estimates.SI evaluation7 min readPDF report
  63. Methods noteVersioning a dataset when its sources changeDataset versioning for data whose sources revise and withdraw records: revision and release, two time axes, time travel retention, version numbers and change records, from published standards.Data services7 min readPDF report
  64. ExplainerWhat a data contract specifies, and how to write oneA data contract states a dataset's schema, quality rules, refresh terms and support before delivery. This explainer walks through the open standard and applies it to a commissioned dataset.Data services7 min readPDF report
  65. Research reportWhat training data costs, from filings and studiesHow much training data costs, from filings and published studies: dated figures for licensed content, human labels and model-made labels, what each covers and what it leaves out.Data services7 min readPDF report
  66. Methods noteWriting annotation guidelines: definitions, worked examples and edge casesHow to write annotation guidelines from published practice: label definitions with inclusion and exclusion rules, worked examples, edge cases, pilot rounds and a versioned change log.Labeling and annotation7 min readPDF report
  67. Methods noteBuilding an obligations register from regulation textA method for reading regulation text into a compliance obligations register: one row per obligation and party, with modality, binding status, exceptions, dates and the cited version.Document and regulation intelligence7 min readPDF report
  68. Methods noteBuying training data: what to check before you license a datasetBefore you buy training data, check its datasheet, the rights back to each source, what the license allows for models, the personal data inside and a measured sample.Data marketplace7 min readPDF report
  69. Methods noteCollecting training data: sources, terms and capture recordsData collection for model training as a method: name the sources, check each one's terms and opt-out signals before capture, and record where and when each item came from.Data services7 min readPDF report
  70. ExplainerDeploying LLMs on Kubernetes: serving, scaling and rollbacksLLM deployment on Kubernetes, layer by layer: model servers, inference-aware routing, scaling on queues and token backlog, scale to zero and rollbacks by weight and revert.Kubernetes platforms7 min readPDF report
  71. Research reportGPU rental price indexes compared: what each one countsA GPU rental price index can count posted rates, executed rentals or surveyed contracts. This report sets each publisher's method and governance side by side, from its own documents.Rillor Compute Index7 min readPDF report
  72. Methods noteHow to specify a training dataset before anyone builds itHow to create a dataset for LLM training starts with a written specification: the use, sources and rights, fields and labels, format, coverage, size, held-out data and acceptance tests.Data services7 min readPDF report
  73. ExplainerLLM nondeterminism: why the same input gets different answersWhy the same request to an LLM can return different answers even at temperature 0: sampling, floating-point order, server load, hardware and serving changes under one model ID.SI evaluation7 min readPDF report
  74. ExplainerMCP security checklist: servers, scopes and pinned tool descriptionsMCP security best practices as one checklist for superintelligence (SI) agents: which servers to allow, which scopes to grant and how to pin and re-approve tool descriptions.Agentic security7 min readPDF report
  75. Methods noteMeasuring label quality with inter-annotator agreementHow percent agreement, Cohen's kappa, Fleiss' kappa and Krippendorff's alpha fit each annotation design, what inter-annotator agreement misses, and the label error rates published test sets carry.Labeling and annotation7 min readPDF report
  76. ExplainerNeocloud and hyperscaler GPU prices, compared like for likeNeoclouds and hyperscalers sell the same H100 in different packages. Posted prices for 8-GPU nodes on 7 November 2026, put on one basis, and what each price includes.GPU and neocloud pricing7 min readPDF report
  77. ExplainerPrompt injection: how it works and how to test an agentWhy text an agent reads can override its instructions, and what prompt injection testing has to measure on superintelligence (SI) agents and the tools they call.Agentic security7 min readPDF report
  78. ExplainerRegulatory change management from official sourcesWhat the Federal Register API, the eCFR, the Unified Agenda, self-regulatory rule filings and legislation.gov.uk publish about a rule change, and how far each can be relied on.Document and regulation intelligence7 min readPDF report
  79. ExplainerSI and Executive Order 14434: what the order changesWhat superintelligence (SI) means under Executive Order 14434, who the order binds, what it leaves as written and how federal documents used the new term after 2 October 2026.Document and regulation intelligence7 min readPDF report
  80. ExplainerStorage for model training: datasets, checkpoints and weightsStorage requirements for model training and checkpointing, from published sizes and measurements: datasets read on every pass, checkpoints written in bursts and weights kept by version.Distributed storage7 min readPDF report
  81. ExplainerThe OWASP Top 10 for Agentic Applications as a test planEach entry of the OWASP Top 10 for Agentic Applications, ASI01 to ASI10, as test cases for superintelligence (SI) agents: the input to plant, the pass condition and the evidence.Agentic security7 min readPDF report
  82. ExplainerWhat a training data license grantsLicensing data for model training, clause by clause: the licensed material, permitted uses, models and outputs, scope, term, warranties, source terms and the records each grant needs.Acquisition and licensing7 min readPDF report
  83. Research reportWhy multi-agent LLM systems fail: what the failure studies showThe MAST taxonomy, debate studies and failure attribution benchmarks show why multi-agent LLM systems fail: loose specifications, dropped correct answers, shallow verification and errors that spread.Agentic systems7 min readPDF report
  84. ExplainerAgent evaluation metrics and what each one tells youHow task success, pass@k and pass^k, tool-call accuracy, escalation, cost per success and time horizon are computed, from the sources that define them, and what each misses.SI evaluation7 min readPDF report
  85. ExplainerBrier score and log loss for probability forecastsThe Brier score and log loss both reward forecasters for stating the probabilities they believe, but they penalize confident misses differently, and the choice can reorder rankings.Forecasting and market intelligence7 min readPDF report
  86. ExplainerCompany disclosure timing and the 23-hour trading dayAfter-hours earnings releases reach markets by wire, website, social post and EDGAR on separate clocks, and Nasdaq's approved 23-hour trading day changes which trade first reflects them.Public communications and markets7 min readPDF report
  87. ExplainerData-center Tier levels: what Tier III and IV coverUptime Institute's four data center Tier levels set maintenance and fault-tolerance criteria with no uptime percentage attached, and NVIDIA says dual feeds give a DGX B200 rack only N.Data-center capacity7 min readPDF report
  88. ExplainerDGX, HGX and MGX explained for GPU system buyersDGX is NVIDIA's complete system, HGX is the GPU baseboard server makers build around, and MGX is a modular reference architecture, so each name changes what a price covers.GPU systems7 min readPDF report
  89. Methods noteDictionaries, FinBERT or language models for financial sentimentHow dictionaries, FinBERT and prompted language models compare in published financial sentiment tests, and what to record so each label can be reproduced after prompts and models change.Public communications and markets7 min readPDF report
  90. ExplainerFPDS, SAM.gov and USAspending: where contract data comes fromContracting officers report contract actions to FPDS, whose search has run only on SAM.gov since February 2026, and USAspending.gov republishes those records with its own delays and gaps.Public procurement records7 min readPDF report
  91. ExplainerHGX H200, B200 and B300 comparedFrom HGX H200 to B300, GPU memory per system grows from 1.1 TB to over 2 TB, while dense FP8 holds at the B200 level and native FP64 falls.GPU systems7 min readPDF report
  92. ExplainerHow entity resolution matches records across sourcesHow entity resolution decides which records describe the same person or company: blocking, the Fellegi-Sunter rule, clustering, registry identifiers, language models and measuring the result.Data services7 min readPDF report
  93. Methods noteHow much GPU memory LLM inference needsLLM inference memory is weights plus a KV cache that grows with context and users, with the arithmetic for a 70B model on 80, 94 and 141 GB GPUs.Private SI cloud7 min readPDF report
  94. Methods noteHow to check LLM citations against their sourcesA step-by-step method to verify each claim in an LLM answer against the passage it cites, with the citation failure rates that published studies report.Agentic systems7 min readPDF report
  95. Methods noteHow to test document extraction accuracy before you buyA pilot plan for LLM document extraction: your own labeled documents, a scoring rule per field, recall where omissions matter, table structure scores and confidence checked against errors.Document and regulation intelligence7 min readPDF report
  96. ExplainerMCP access control: scoped credentials, allow-lists and audit logsHow the Model Context Protocol connects superintelligence (SI) assistants to business tools, and the controls that keep that access governed: tenant isolation, scoped credentials, allow-lists and audit logs.Agentic security5 min readPDF report
  97. Methods noteReading GPU prices from cloud pricing APIsThe GPU pricing APIs of AWS, Azure, Google Cloud, Oracle and Lambda differ in access, units and coverage, so each needs its own conversion to dollars per GPU-hour.GPU and neocloud pricing7 min readPDF report
  98. Methods noteRebuilding an order book from snapshots and diffsOrder book reconstruction applies sequenced diffs to a full snapshot under each venue's own numbering and checks, and one missed update leaves the book wrong until it is resynchronized.Market data7 min readPDF report
  99. Methods noteRent or buy GPUs: a break-even method from public inputsOwning GPUs costs less than renting on demand only above a break-even utilization, and against committed terms the test is cost per reserved hour, built from dated public inputs.Rillor Compute Index7 min readPDF report
  100. Methods noteSelf-hosted LLM vs API cost: a break-even methodA self-hosted model costs less than an API only above a break-even utilization, set by GPU hourly cost, tokens served within a latency limit and the API price.Private SI cloud7 min readPDF report
  101. ExplainerTAQ data explained: timestamps, sale conditions and round lotsNYSE's Daily TAQ files hold every trade and quote on the U.S. consolidated tapes, and their timestamp, sale condition and round lot rules changed on set dates.Market data7 min readPDF report
  102. Research reportWhat agent benchmarks measure, and what they missHow SWE-bench, GAIA, WebArena, OSWorld and tau-bench score LLM agents, and the validity and contamination flaws that audits have found.SI evaluation7 min readPDF report
  103. ExplainerWhat an HGX B200 or B300 server price includesAn eight-GPU HGX server price covers the GPU baseboard, host CPUs, memory, network cards, storage, power supplies and warranty, and dated listings show what memory, network cards and CPUs add.GPU systems pricing7 min readPDF report
  104. Research reportWhat benchmarks show about LLM contract review accuracyEight benchmarks from 2021 to 2026 test LLM contract review, and scores run high on yes-or-no clause questions and lower on clause extraction and final review.Document and regulation intelligence7 min readPDF report
  105. ExplainerWhat makes a GPU capacity contract financeableThe terms lenders test in a GPU offtake agreement before they lend against it, read from GPU capacity contracts and credit agreements filed with the SEC.GPU financing7 min readPDF report
  106. Research reportWhat research shows about training on spot GPUsSystems research and provider documents show what spot GPU instances save on model training once lost work, checkpoint overhead and deadlines are counted.GPU and neocloud pricing7 min readPDF report
  107. Research reportNovember 2025 to November 2026What sets GPU system lead times in 2026Filings and earnings calls show what sets a GPU server lead time: packaging, high-bandwidth memory, substrates and testing, system builders and powered buildings.Forward agreements7 min readPDF report
  108. ExplainerWhat to test before an agent buys on your behalfPayment protocols for agents record what a user authorized, and published research shows which agentic commerce risks remain inside those limits and how to test each.Market simulation7 min readPDF report
  109. ExplainerWhen a multi-agent system beats a single agentControlled studies of single agent vs multi-agent designs find extra agents help when work splits into independent parts and hurt when each step depends on the last.Agentic systems7 min readPDF report
  110. ExplainerForward agreements for GPU systems: how they workA forward purchase agreement fixes the grade, quantity, price and delivery month of GPU systems at signing, with physical delivery later and no exchange or clearinghouse involved.Forward agreements7 min readPDF report
  111. Explainer2025 to 2026Perpetual DEX market structure, 2025 to 2026How on-chain perpetual futures venues match orders, charge fees, set funding and oracle prices, and what public data shows about their volume, open interest and real-world-asset markets.Market data5 min readPDF report
  112. Research report2024 to 2026LLM agents as market participants: what the research showsA review of published studies that put LLM agents in auctions, pricing, negotiation, marketplaces and trading: the biases observed, the collusion and deception findings, and the evaluation gaps.Market simulation6 min readPDF report
  113. ExplainerGPU financing structures explainedHow leases, loans, sale-leasebacks, vendor and milestone financing, and contract-backed facilities pay for GPU systems, and what lenders check before they commit.GPU financing5 min readPDF report
  114. Methods noteMeasuring LLM agents in simulated auctionsHow to test LLM agents as bidders: induced values, known efficient outcomes, measures read from bids alone and human reference agents built from published experiments.Market simulation6 min readPDF report
  115. Market reviewJuly to September 2026Compute market review: Q3 2026Looking back at July to September 2026: buyers raised spending while short of capacity, newer GPUs rented for more, and compute futures went under CFTC review.Rillor Compute Index5 min readPDF report
  116. Market reviewApril to June 2026Compute market review: Q2 2026Looking back at April to June 2026: capital spending plans rose, memory became the cost to watch, H100 rents split by tier and exchanges announced GPU futures plans.Rillor Compute Index5 min readPDF report
  117. Market reviewJanuary to March 2026Compute market review: Q1 2026Looking back at January to March 2026: GPU rental capacity sold out across generations, and one-year H100 contract prices rose almost 40% from their October low.Rillor Compute Index6 min readPDF report
  118. Market reviewOctober to December 2025Compute market review: Q4 2025Looking back at October to December 2025: buyers committed to GB300 systems at scale while powered buildings and grid capacity set the pace of deployment.Rillor Compute Index5 min readPDF report
  119. Market reviewJuly to September 2025Compute market review: Q3 2025Looking back at July to September 2025: posted H100 rents drifted lower while buyers signed multi-year capacity contracts worth tens of billions of dollars.Rillor Compute Index5 min readPDF report
  120. Research reportCross-venue price discovery: regulated crude futures and on-chain perpetualsRillor tested whether CME WTI futures lead an on-chain crude perpetual. The lead held out of sample. No net edge after execution costs reached statistical significance.Forecasting and market intelligence7 min readPDF report
  121. Methods noteEvidence classes in GPU pricingEvidence about a GPU system's price comes in six classes, from specification to verified transaction, and mixing them misleads valuations and indexes.GPU systems pricing6 min readPDF report
  122. Methods noteFrom rule text to a checkable modelA worked example on SEC Rule 201: fields extracted by an LLM and tied to exact phrases, reviewed by a person, then checked against a stated property.Document and regulation intelligence7 min readPDF report
  123. Research reportJuly 2023 to October 2026Neocloud economics: contract first, hardware secondFilings show GPU cloud operators sign customers before buying hardware, then borrow against those contracts at interest rates set by the customer's credit.GPU financing5 min readPDF report

Compute market review

Each review covers one period of the market for SI compute. They are listed in order of the period covered.

  1. July to September 2025

    Compute market review: Q3 2025

    Read: Compute market review: Q3 2025Full report: Compute market review: Q3 2025
  2. October to December 2025

    Compute market review: Q4 2025

    Read: Compute market review: Q4 2025Full report: Compute market review: Q4 2025
  3. January to March 2026

    Compute market review: Q1 2026

    Read: Compute market review: Q1 2026Full report: Compute market review: Q1 2026
  4. April to June 2026

    Compute market review: Q2 2026

    Read: Compute market review: Q2 2026Full report: Compute market review: Q2 2026
  5. July to September 2026

    Compute market review: Q3 2026

    Read: Compute market review: Q3 2026Full report: Compute market review: Q3 2026

Agent evaluation

Listed in order of publication.

  1. 2024 to 2026

    LLM agents as market participants: what the research shows

    Read: LLM agents as market participants: what the research showsFull report: LLM agents as market participants: what the research shows
  2. 9 October 2026

    Measuring LLM agents in simulated auctions

    Read: Measuring LLM agents in simulated auctionsFull report: Measuring LLM agents in simulated auctions

Topics

One page for each field we work in, with its research and how to work with us on it.

Notices

Rillor is not a registered investment adviser or commodity trading advisor and does not provide investment or trading advice. Research, forecasts, data and software described on this site are for research and engineering use. Nothing here is an offer or recommendation to buy or sell any security, commodity interest or digital asset. Past or simulated results do not indicate future results.

NVIDIA, AMD, Ceph, Kubernetes and related product names are trademarks of their owners and are used only to identify products. No affiliation or endorsement is implied.

Ask about the data behind a report.

Talk to usHow we work