1% ============================================================================2% Part VI — Systems and Computing Infrastructure3% Written by Agent F (Systems Specialist), WP44, 2026-11-03.4% Sources: literature_synthesis/agent_F/{review.md,claims.csv,contributions.csv,open_problems.csv};5% bibliography: references.bib (group J and systems canon) + references_part06.bib (verified; references_verification_part06.csv).6% Corpus quantities only through the macros of tables/numbers.tex (prefixes Cagr*, Catcount*, ItsSeriesCat*, Sysflow*, Selection*).7% ============================================================================89\section{Scope and evidence}10\label{sec:p6-scope}1112This part reviews the eleven categories of the taxonomy that describe how computation is built, housed and delivered: computer architecture, accelerators and AI hardware, operating systems, distributed computing, cloud computing, edge computing, computer networks, database systems, information retrieval, quantum computing and sustainable computing. In the area partition of Appendix~A the first six belong to \emph{data and distributed systems} and the last five to \emph{hardware and emerging paradigms}; the grouping adopted here follows the intellectual dependencies instead, so that accelerators are read with architecture, operating systems with distributed systems, and information retrieval with databases.1314The evidence is of two kinds and the two are kept apart throughout. The first is the corpus itself: the yearly primary-category counts and growth rates of Part~IX (Tables~\ref{tab:cagr} and~\ref{tab:systems_categories}), the structural-change estimates of the monthly log-volume series (Table~\ref{tab:breaks}), and the citation flows between categories of the knowledge graph. The second is the landmark lists of the two areas, \SelectionByAreaDataDistributed{} records for data and distributed systems and \SelectionByAreaHardwareEmerging{} for hardware and emerging paradigms, selected by the rule of Part~III (the most cited works of each primary category under the field-normalised thirty-six-month window, with a floor of twenty-five records per category; for 2025 and 2026, twelve-month citations relative to the year mean). What is known about these works is their metadata and, for most of them, their abstract; no full text was consulted. Consequently every statement about the content of a paper below reports what its abstract says, and a paper whose abstract is absent from the corpus is described from its title and venue only. Preprints are named as preprints. The structured record of each claim, with its evidence type and confidence, is in the file \texttt{literature\_synthesis/agent\_F/claims.csv}.1516Two properties of the landmark lists bound what they can show. The selection rule favours records that OpenAlex has dated to a re-publication rather than to the original: the distributed-computing list opens with MapReduce \citep{dean2008mapreduce}, Lamport's ``Time, clocks'' and ``The Byzantine generals problem'' recorded as 2018--2019 items, and the computer-architecture list is headed by the NumPy paper. Such records measure what the field cites, not what it produced, and they are set aside. The rule also admits, for 2025--2026, records with very few citations in absolute terms, so that the most recent ``landmarks'' are indications of current citation rather than established contributions. More consequential is classification noise. Roughly one record in five in the two lists is not systems research at all: dissertations on language teaching and admission systems under cloud computing and operating systems, an ethnography of mobile phones and the Paleobiology Database user guide under database systems, nursing students' AI literacy and regulatory sandboxes under operating systems, mammography screening under quantum computing, higher-education surveys under sustainable computing, control-theory papers under distributed computing, molecular-dynamics codes under computer architecture. These records are listed, with their identifiers, in \texttt{literature\_synthesis/agent\_F/review.md}; they are excluded from the synthesis, and their frequency is reported because it bounds the precision of every category-level statistic that follows. The taxonomy's validation in Appendix~A did not sample these categories at the depth that would estimate the error rate, and this is recorded as an open problem for the data-quality discussion of Part~XI.1718\section{The systems categories in the corpus}19\label{sec:p6-corpus}2021\subsection{Volume, growth and share}22\label{sec:p6-growth}2324Table~\ref{tab:systems_categories} assembles, for the eleven categories, the primary-category output in 2018, 2022 and 2025, the compound annual growth rates over the baseline (2018--2022) and post (2023--2025) periods, and the structural-change statistics discussed below. The first regularity is a division into two groups. Seven classical categories declined in absolute output between 2018 and 2022 and grew between 2023 and 2025: computer architecture (\CagrComputerArchitectureCagrBaseline{} then \CagrComputerArchitectureCagrPost{} per year), operating systems (\CagrOperatingSystemsCagrBaseline{} then \CagrOperatingSystemsCagrPost), distributed computing (\CagrDistributedComputingCagrBaseline{} then \CagrDistributedComputingCagrPost), cloud computing (\CagrCloudComputingCagrBaseline{} then \CagrCloudComputingCagrPost), computer networks (\CagrComputerNetworksCagrBaseline{} then \CagrComputerNetworksCagrPost), database systems (\CagrDatabaseSystemsCagrBaseline{} then \CagrDatabaseSystemsCagrPost) and information retrieval (\CagrInformationRetrievalCagrBaseline{} then \CagrInformationRetrievalCagrPost). Four categories grew in both periods: edge computing (\CagrEdgeComputingCagrBaseline{} then \CagrEdgeComputingCagrPost), quantum computing (\CagrQuantumComputingCagrBaseline{} then \CagrQuantumComputingCagrPost), sustainable computing (\CagrSustainableComputingCagrBaseline{} then \CagrSustainableComputingCagrPost) and AI accelerators (\CagrAiAcceleratorsCagrBaseline{} then \CagrAiAcceleratorsCagrPost). The 2026 counts are excluded from the growth rates because the year is incomplete.2526The second regularity concerns shares rather than levels. Computer networks, the largest of the eleven, held \CagrComputerNetworksShareTwoZeroOneEight{} of all primary output in 2018 and \CagrComputerNetworksShareTwoZeroTwoFive{} in 2025; computer architecture fell from \CagrComputerArchitectureShareTwoZeroOneEight{} to \CagrComputerArchitectureShareTwoZeroTwoFive, database systems from \CagrDatabaseSystemsShareTwoZeroOneEight{} to \CagrDatabaseSystemsShareTwoZeroTwoFive, distributed computing from \CagrDistributedComputingShareTwoZeroOneEight{} to \CagrDistributedComputingShareTwoZeroTwoFive, cloud computing from \CagrCloudComputingShareTwoZeroOneEight{} to \CagrCloudComputingShareTwoZeroTwoFive{} and operating systems from \CagrOperatingSystemsShareTwoZeroOneEight{} to \CagrOperatingSystemsShareTwoZeroTwoFive. Quantum computing rose from \CagrQuantumComputingShareTwoZeroOneEight{} to \CagrQuantumComputingShareTwoZeroTwoFive{} and sustainable computing from \CagrSustainableComputingShareTwoZeroOneEight{} to \CagrSustainableComputingShareTwoZeroTwoFive. The relative decline of the classical systems fields is therefore a phenomenon of the whole period: it began before November 2022 and continued after it, while the absolute output of those fields recovered. Any reading of the post-2022 recovery as a response to a new workload must be weighed against the simpler possibility of reversion after a four-year decline whose causes this corpus does not identify.2728\subsection{Structural change}29\label{sec:p6-breaks}3031The interrupted time-series model of Part~IV, fitted to the monthly log-volume series of each category with the event at December 2022, yields a positive change in trend for all eleven categories, with heteroskedasticity-and-autocorrelation-consistent $p$-values below one per cent in every case. The change is smallest for edge computing (\ItsSeriesCatEdgeComputingItsChatgptSlopeChange{} per month, $p = \ItsSeriesCatEdgeComputingItsChatgptSlopeP$) and largest for information retrieval (\ItsSeriesCatInformationRetrievalItsChatgptSlopeChange), with distributed computing (\ItsSeriesCatDistributedComputingItsChatgptSlopeChange), cloud computing (\ItsSeriesCatCloudComputingItsChatgptSlopeChange), accelerators (\ItsSeriesCatAiAcceleratorsItsChatgptSlopeChange), quantum computing (\ItsSeriesCatQuantumComputingItsChatgptSlopeChange), architecture (\ItsSeriesCatComputerArchitectureItsChatgptSlopeChange), operating systems (\ItsSeriesCatOperatingSystemsItsChatgptSlopeChange), sustainable computing (\ItsSeriesCatSustainableComputingItsChatgptSlopeChange), database systems (\ItsSeriesCatDatabaseSystemsItsChatgptSlopeChange) and networks (\ItsSeriesCatComputerNetworksItsChatgptSlopeChange) in between. The uniformity is the finding. A steepening of the same order in quantum computing, which has no methodological connection with language models, and in information retrieval, which has a direct one, is not the signature of a workload-specific shock; it is the signature of a field-wide acceleration whose timing the placebo analysis then tests.3233The placebo test rejects date-specificity for every systems category. The share of alternative event dates whose slope statistic is at least as extreme as the one estimated at December 2022 would have to be small for the date to be singled out, and it is never small: \ItsSeriesCatComputerArchitecturePlaceboChatgptSlopeTPlaceboP{} for architecture, \ItsSeriesCatDistributedComputingPlaceboChatgptSlopeTPlaceboP{} for distributed computing, \ItsSeriesCatEdgeComputingPlaceboChatgptSlopeTPlaceboP{} for edge computing, \ItsSeriesCatDatabaseSystemsPlaceboChatgptSlopeTPlaceboP{} for databases, \ItsSeriesCatInformationRetrievalPlaceboChatgptSlopeTPlaceboP{} for information retrieval, \ItsSeriesCatQuantumComputingPlaceboChatgptSlopeTPlaceboP{} for quantum computing, \ItsSeriesCatAiAcceleratorsPlaceboChatgptSlopeTPlaceboP{} for accelerators, \ItsSeriesCatSustainableComputingPlaceboChatgptSlopeTPlaceboP{} for sustainable computing, \ItsSeriesCatComputerNetworksPlaceboChatgptSlopeTPlaceboP{} for networks, \ItsSeriesCatCloudComputingPlaceboChatgptSlopeTPlaceboP{} for cloud computing and \ItsSeriesCatOperatingSystemsPlaceboChatgptSlopeTPlaceboP{} for operating systems. Level changes at the event month are heterogeneous in sign: negative and significant at the five-per-cent level for information retrieval (\ItsSeriesCatInformationRetrievalItsChatgptLevelChange, $p = \ItsSeriesCatInformationRetrievalItsChatgptLevelP$), accelerators (\ItsSeriesCatAiAcceleratorsItsChatgptLevelChange, $p = \ItsSeriesCatAiAcceleratorsItsChatgptLevelP$) and quantum computing (\ItsSeriesCatQuantumComputingItsChatgptLevelChange, $p = \ItsSeriesCatQuantumComputingItsChatgptLevelP$); positive and significant for computer networks (\ItsSeriesCatComputerNetworksItsChatgptLevelChange, $p = \ItsSeriesCatComputerNetworksItsChatgptLevelP$) and sustainable computing (\ItsSeriesCatSustainableComputingItsChatgptLevelChange, $p = \ItsSeriesCatSustainableComputingItsChatgptLevelP$); indistinguishable from zero for the other six, with cloud computing the clearest null (\ItsSeriesCatCloudComputingItsChatgptLevelChange, $p = \ItsSeriesCatCloudComputingItsChatgptLevelP$). The placebo $p$-values for the level statistics exceed \ItsSeriesCatAiAcceleratorsPlaceboChatgptLevelTPlaceboP{} in every category. The Bai--Perron break dates selected by the Bayesian information criterion, reported in Table~\ref{tab:systems_categories}, are scattered from the beginning of 2022 to the spring of 2025; cloud computing is the only category with a selected break in the month of the ChatGPT release, and the single break of the networks series precedes it. The reading that the corpus supports is a broad acceleration of systems output from 2022--2023 onward that the November 2022 date does not single out, superimposed on the prior decline of the classical categories.3435\subsection{Citation flows}36\label{sec:p6-flows}3738The category-to-category citation flows of the knowledge graph, aggregated over citing years 2018--2022 and 2023--2026,\footnote{Calendar citing years; the aggregation does not use the 30 November 2022 boundary. Source: \texttt{results/graph/category\_flows\_yearly.csv}, aggregated by the macros of \texttt{tables/make\_numbers.py}.} show the systems categories to be strongly self-referential and stable. In the post years the share of references that a category's works send to works of the same category is \SysflowQuantumComputingPostToQuantumComputingShare\% for quantum computing, \SysflowComputerNetworksPostToComputerNetworksShare\% for computer networks, \SysflowInformationRetrievalPostToInformationRetrievalShare\% for information retrieval, \SysflowComputerArchitecturePostToComputerArchitectureShare\% for computer architecture, \SysflowDatabaseSystemsPostToDatabaseSystemsShare\% for database systems, \SysflowAiAcceleratorsPostToAiAcceleratorsShare\% for accelerators, \SysflowCloudComputingPostToCloudComputingShare\% for cloud computing, \SysflowDistributedComputingPostToDistributedComputingShare\% for distributed computing, \SysflowEdgeComputingPostToEdgeComputingShare\% for edge computing, \SysflowOperatingSystemsPostToOperatingSystemsShare\% for operating systems and \SysflowSustainableComputingPostToSustainableComputingShare\% for sustainable computing. None of these shares moved by more than a few points from the baseline years (operating systems from \SysflowOperatingSystemsBaselineToOperatingSystemsShare\%, quantum computing from \SysflowQuantumComputingBaselineToQuantumComputingShare\%). Three details are relevant to the synthesis. The large-language-model category receives \SysflowInformationRetrievalPostToLlmShare\% of information-retrieval references in the post years against \SysflowInformationRetrievalBaselineToLlmShare\% before, the only systems category in which that destination becomes measurable. Operating systems now sends more references to computer architecture (\SysflowOperatingSystemsPostToComputerArchitectureShare\%) than to itself, which is consistent with the GPU-centred systems work reviewed in Section~\ref{sec:p6-os}. Edge computing and sustainable computing cite computer networks (\SysflowEdgeComputingPostToComputerNetworksShare\% and \SysflowSustainableComputingPostToComputerNetworksShare\%) more than their own category in both periods, a fact about where the classifier draws its boundaries as much as about intellectual dependence. The flows give no sign of a reorganisation of the systems fields around artificial intelligence between the two periods; what changed, as the following sections show, is the content of the most cited works within categories whose citation neighbourhoods stayed where they were.3940\section{Computer architecture and accelerators}41\label{sec:p6-arch}4243\subsection{The end of scaling and the arrival of one workload}44\label{sec:p6-arch-scaling}4546The architecture literature of the baseline period is organised around a diagnosis that predates it: the end of Dennard scaling \citep{dennard1974scaling} and the slowing of Moore's law left domain-specific architecture as the remaining route to performance, a position stated in the Turing lecture of \citet{hennessy2019goldenage} and quantified by \citet{leiserson2020roomattop}, who locate the remaining gains in software, algorithms and hardware specialisation rather than in transistors. The tensor processing unit \citep{jouppi2017tpu} and the GPU lineage \citep{dally2021gpu} are the canonical instances; \citet{sze2017efficientdnn} is the standard account of why neural-network inference rewards specialised data paths. The landmark lists confirm that this is the lineage the field cites, and they add the processing-in-memory primer of \citet{mutlu2023pim} and the concurrency analysis of distributed deep learning by \citet{bennun2020demystifying} as the most cited expository works of the pre-2022 years. What the baseline lists also contain, and what is set aside, are scientific-software papers (NumPy, NAMD, Quantum ESPRESSO) that the classifier assigns to architecture because their topics concern CPU and GPU execution.4748After 2022 the most cited architecture works, once misclassified records are removed, form a single set: systems papers for the training and serving of large language models. This is the central empirical observation of this section and it is a statement about metadata, not about the field's full output: the citation-ranked top of the category is occupied by one workload. The corpus statistics of Section~\ref{sec:p6-corpus} qualify it at once. Architecture's trend change at December 2022 (\ItsSeriesCatComputerArchitectureItsChatgptSlopeChange{} per month) is in the middle of the systems range, its level change is null ($p = \ItsSeriesCatComputerArchitectureItsChatgptLevelP$), and its share of computer-science output fell from \CagrComputerArchitectureShareTwoZeroOneEight{} to \CagrComputerArchitectureShareTwoZeroTwoFive. The concentration of citation on LLM systems papers occurred within a category that did not grow faster than the rest and whose relative weight continued to decline.4950\subsection{Training and inference systems}51\label{sec:p6-arch-systems}5253The training-systems line is largely a pre-window inheritance recorded in preprints and conference papers: tensor and pipeline parallelism in Megatron-LM \citep[preprint]{shoeybi2019megatronlm} and its scaling to thousands of GPUs \citep{narayanan2021megatron}, memory partitioning of optimiser state in ZeRO \citep{rajbhandari2020zero}, fully sharded data parallelism in PyTorch \citep[preprint]{zhao2023fsdp}, and the asynchronous dataflow of Pathways \citep[preprint]{barham2022pathways}. The compiler stack beneath these systems is the one documented in Part~VII: LLVM and MLIR \citep{lattner2004llvm,lattner2021mlir}, TVM \citep[preprint]{chen2018tvm}, Triton \citep{tillet2019triton}. The landmark set adds the paper that brought graph compilation into the dominant framework: \citet{ansel2024pytorch2} describe TorchDynamo, which captures computation graphs by rewriting Python bytecode before execution, and TorchInductor, which lowers them to Triton for GPUs and to C++ for CPUs; the abstract reports geometric-mean speedups of 2.27$\times$ for inference and 1.41$\times$ for training over eager execution on the authors' model suite. This is a vendor-authored evaluation of the authors' own system, which is the norm in this literature and the reason the claims below are reported as the authors' rather than as established.5455The inference-systems line is where the window's own contributions lie, and it is organised by one bottleneck. \citet{gholami2024memorywall}, in a perspective article in \emph{IEEE Micro}, state that peak server FLOPS grew at $3.0\times$ per two years over the past twenty years while DRAM bandwidth grew at $1.6\times$ and interconnect bandwidth at $1.4\times$, and argue that memory bandwidth, not arithmetic throughput, has become the binding constraint for decoder-only Transformers at serving time. The trend figures are the authors' and are not re-derived here, but the diagnosis organises the rest of the literature. \citet{pope2023inference}, in a preprint, derive the latency--throughput frontier of Transformer inference under memory and communication limits; FlashAttention \citep[preprint]{dao2022flashattention} reduces memory traffic of the attention kernel by tiling; quantisation methods reduce the weight bytes moved per token, from post-training GPTQ \citep[preprint]{frantar2023gptq} to activation-aware AWQ, whose abstract reports an algorithm--system co-design with an on-device framework (TinyChat) that converts memory savings into speedups through on-the-fly dequantisation and kernel fusion \citep{lin2025awq}. The serving systems proper are discussed with operating systems in Section~\ref{sec:p6-os}, because their abstraction is memory management; here it suffices to note that Splitwise \citep{patel2024splitwise}, an ISCA paper, characterises each inference request as a compute-intensive prompt phase and a memory-intensive token-generation phase with distinct latency, throughput and power profiles, and places the two phases on separate machine pools so that generation runs on older or lower-power hardware. Taken together these works establish, at the level of abstracts, that the architecture community's response to the LLM workload in 2023--2025 was to redesign the memory system and the deployment topology rather than the arithmetic unit. Whether the memory-wall diagnosis survives changes in the installed base, and whether gains attributed to it come from bandwidth or from model-side sparsity, are open (Section~\ref{sec:p6-synthesis}).5657The survey of \citet{silvano2025accelerators} closes the period: it classifies accelerators for heterogeneous HPC platforms into GPU and TPU systems, FPGA and ASIC designs, neural processing units, open RISC-V accelerators, and emerging-memory and neuromorphic devices. That such a survey, and the earlier benchmarking of commercial accelerators by \citet{reuther2022accelerators} and TPU v4 \citep{jouppi2023tpuv4}, do not appear in the accelerator landmark list but in the bibliography is a reminder that the classifier routes digital accelerators to architecture.5859\subsection{Photonic, neuromorphic and in-memory accelerators}60\label{sec:p6-accel}6162The accelerator category is the one most affected by classification: at least half of its post-period landmarks are memristive and chaotic Hopfield-network circuit papers in journals of nonlinear dynamics and circuit design, a literature on attractors, multistability and image encryption that the topic classifier maps to AI hardware. Removing it leaves a photonic and neuromorphic thread whose baseline anchors are Loihi, Intel's neuromorphic manycore processor with programmable synaptic learning rules \citep{davies2018loihi}, and the perspective on neuromorphic algorithms and applications of \citet{schuman2022neuromorphic}. The post-period additions are optical: the Taichi photonic chiplet \citep{xu2024taichi}, whose abstract reports a diffractive--interference hybrid design with a distributed architecture reaching 160~TOPS/W and on-chip classification over a thousand categories; the universal photonic accelerator of \citet{ahmed2025photonic}, for which the corpus holds no abstract; a review of optical neural networks \citep{fu2024onn}; and, in 2026, a monolithically integrated neuromorphic photonic circuit with on-chip capacitive analog memory, whose abstract reports an analysed power saving of more than 26$\times$ relative to SRAM--DAC designs \citep{lam2026neuromorphicphotonic}, together with a hyper-multiplexed matrix--matrix optical processor.6364These are device-level demonstrations evaluated on small tasks, and the efficiency figures they report are figures for the photonic core. None of the abstracts reports end-to-end energy including digital-to-analog and analog-to-digital conversion, memory traffic and precision loss, and none compares against a precision-matched digital accelerator under a shared benchmark. The literature is therefore classified here as speculative with respect to its claim on the computing infrastructure: the physics is established, the system-level advantage is not. The corpus statistics reinforce the caution. The accelerator category's level fell at December 2022 (\ItsSeriesCatAiAcceleratorsItsChatgptLevelChange, $p = \ItsSeriesCatAiAcceleratorsItsChatgptLevelP$), its post-period growth (\CagrAiAcceleratorsCagrPost{} per year) is slower than its baseline growth (\CagrAiAcceleratorsCagrBaseline), and its references go to deep-learning architectures (\SysflowAiAcceleratorsPostToDeepLearningArchitecturesShare\%) and computer architecture (\SysflowAiAcceleratorsPostToComputerArchitectureShare\%) in proportions that did not change between periods. Whatever occurred in AI hardware after 2022 is not visible as a surge in this category's volume or as a reorientation of its citations.6566\section{Operating systems and distributed systems}67\label{sec:p6-os}6869\subsection{Operating systems: memory abstractions for accelerators}70\label{sec:p6-os-os}7172Operating systems is the smallest of the eleven categories (\CagrOperatingSystemsNTwoZeroTwoFive{} primary works in 2025) and the one that contracted fastest before the window (\CagrOperatingSystemsCagrBaseline{} per year), recovering at \CagrOperatingSystemsCagrPost{} per year in 2023--2025 while its share of computer science continued to fall. Its baseline landmarks are what one would expect of the late 2010s: customisable trusted execution environments on RISC-V in Keystone \citep{lee2020keystone}, sub-millisecond serverless runtimes for interactive microservices in Nightcore \citep{jia2021nightcore}, the Berkeley position paper on serverless computing \citep[preprint]{jonas2019serverless}, and key-value stores for persistent memory. Unikernels \citep{madhavapeddy2013unikernels,kuenzer2021unikraft} and kernel-bypass or in-kernel packet processing with XDP and eBPF \citep{hoiland2018xdp,vieira2020ebpf} belong to the same lineage and are cited by the corpus, though they do not reach the landmark threshold.7374The post-period landmarks are of a different kind. PowerInfer \citep{song2024powerinfer}, an SOSP paper, observes that neuron activations in large models follow a power-law distribution, keeps the consistently active ``hot'' neurons on a consumer GPU and computes the input-dependent ``cold'' ones on the CPU, with predictors and sparse operators to bridge the two. Orion \citep{strati2024orion} intercepts the kernel launches of several clients sharing a GPU and schedules them at the granularity of individual operators so as to limit interference at the scale of tens of microseconds. vAttention \citep{prabhu2025vattention} contests the memory abstraction of vLLM \citep{kwon2023vllm}: where PagedAttention allocates the key--value cache on demand in non-contiguous pages, inspired by virtual memory and paging, vAttention argues that the loss of contiguity imposes programming and performance overheads and restores a contiguous virtual layout through CUDA virtual-memory primitives. This is a live design controversy, and the corpus offers no way to adjudicate it: each system is evaluated by its authors on particular GPU generations and traces, and no neutral, versioned serving benchmark exists. What the abstracts establish is narrower and more interesting: the operating-systems problem of 2023--2025, as the field's most cited work defines it, is the management of accelerator memory and accelerator time for one workload, and the classical concerns of the baseline years---isolation, serverless latency, persistent memory---no longer appear at the top of the category. The citation flows agree: operating-systems works now send more references to computer architecture (\SysflowOperatingSystemsPostToComputerArchitectureShare\%) than to their own category (\SysflowOperatingSystemsPostToOperatingSystemsShare\%).7576Two further items qualify the picture. The ``LLM as operating system'' proposals that the bibliography records \citep[all preprints]{packer2023memgpt,ge2023llmos,mei2024aios} treat context management and tool scheduling for language-model agents as an operating-systems problem by analogy; they are position statements, not systems in the sense of the papers above, and none reaches the landmark list. The standardisation of the BPF instruction set as an IETF RFC \citep{ietf2024rfc9669} in 2024, together with an in-kernel aggregation preprint whose record is misdated in the corpus, suggests that kernel-extension technology remained an infrastructure theme of the period; the evidence is one RFC and one abstract, and the suggestion is recorded as such.7778\subsection{Distributed computing: a canon cited, not renewed}79\label{sec:p6-os-dist}8081The distributed-computing landmark list is dominated by what the field cites rather than by what it produced: re-dated records of MapReduce \citep{dean2008mapreduce}, Lamport's logical clocks and the Byzantine generals problem head the baseline, and the canon that the bibliography documents---Paxos \citep{lamport1998paxos}, practical Byzantine fault tolerance \citep{castro2002pbft}, the CAP formulation \citep{gilbert2002cap}, Spanner \citep{corbett2013spanner}, Spark \citep{zaharia2016spark}, Borg and Kubernetes \citep{verma2015borg,burns2016kubernetes}---is present in the references of the period without any post-2022 work of comparable standing in consensus, transactions or fault tolerance reaching the list. The growth topic of the baseline years was blockchain: a blockchain survey appears twice under distinct OpenAlex identifiers, the selfish-mining paper ``Majority is not enough'' and a consensus pre-processing method are among the most cited, and in the post years distributed-computing works still send \SysflowDistributedComputingPostToBlockchainShare\% of their references to the blockchain category. The category's trend change at December 2022 is among the largest of the eleven (\ItsSeriesCatDistributedComputingItsChatgptSlopeChange{} per month) and its output recovered at \CagrDistributedComputingCagrPost{} per year after declining at \CagrDistributedComputingCagrBaseline; two Bai--Perron breaks are selected, neither in the event month (Table~\ref{tab:systems_categories}).8283The post-period landmarks that survive scrutiny are three. vLLM \citep{kwon2023vllm}, classified here rather than under operating systems, reports that near-zero waste of key--value-cache memory and the sharing of cache blocks within and across requests raise throughput by two to four times at equal latency relative to the previous state of the art. OrbitalBrain proposes in-space distributed training on Earth-observation nanosatellites under bandwidth and power constraints, and a live race-bypassing system for deployed multithreaded applications appears as a 2026 repository record of older work. The first is the window's single most influential systems artefact in this area; the other two are indications of citation rather than landmarks. Federated learning, which the taxonomy treats as a category of its own but which is a distributed-systems topic in its mechanics \citep{mcmahan2017fl,kairouz2021fl,li2020flchallenges}, appears in the systems lists only through surveys, discussed in Section~\ref{sec:p6-sustain}. The conclusion for distributed computing is therefore negative in a precise sense: the corpus records no renewal of the field's theoretical or protocol canon in 2023--2026, and records its most cited new work as an inference-serving system whose intellectual home is the operating-systems tradition of memory management.8485\section{Cloud and edge computing}86\label{sec:p6-cloud}8788\subsection{Cloud computing: the same problem, a new workload}89\label{sec:p6-cloud-cloud}9091Cloud computing's baseline landmarks are scheduling and placement papers---virtual-machine allocation with usage prediction, QoS-aware scheduling for energy conservation, metaheuristic task scheduling---together with surveys of data-centre energy efficiency and of IoT--cloud integration. The post-period landmarks keep the problem and change the workload. The review of deep-reinforcement-learning methods for cloud resource scheduling by \citet{zhou2024drlcloud} is the bridge; the 2026 slice consists almost entirely of scheduling for large-language-model inference: token-level pooling and slicing of GPUs for co-located models, contextual-bandit elastic scheduling for serverless multi-tenant inference \citep{li2026deepservecb}, and carbon-aware prompt scheduling, whose abstract reports, in trace-driven simulation, a reduction of 26.8\% in carbon per thousand generated tokens relative to round-robin routing \citep{liao2026caps}. These 2026 records are provisional landmarks in the sense of Section~\ref{sec:p6-scope}, they appear in smaller IEEE conferences, and each is evaluated in simulation or on a single testbed; they are cited here not as results but as evidence of what the cloud-computing literature of 2026 is about. The one post-period landmark with an operations rather than a scheduling focus is RCACopilot \citep{chen2024rcacopilot}, a EuroSys paper that matches cloud incidents to handlers, aggregates runtime diagnostics and uses a language model to predict the root-cause category and produce an explanatory narrative; the abstract reports accuracy up to 0.766 on a year of incidents from one provider, a proprietary evaluation that no public benchmark allows to be reproduced.9293The corpus statistics for cloud computing are distinctive in one respect. Its level change at December 2022 is the clearest null of the eleven (\ItsSeriesCatCloudComputingItsChatgptLevelChange, $p = \ItsSeriesCatCloudComputingItsChatgptLevelP$), its trend change is \ItsSeriesCatCloudComputingItsChatgptSlopeChange{} per month, and it is the only systems category for which the Bayesian information criterion selects a Bai--Perron break in the month of the ChatGPT release, with a second break two years later. The placebo result (\ItsSeriesCatCloudComputingPlaceboChatgptSlopeTPlaceboP) does not permit the coincidence to be read as specific, and the category's output had been contracting at \CagrCloudComputingCagrBaseline{} per year before recovering at \CagrCloudComputingCagrPost. The serverless agenda of the baseline \citep{jonas2019serverless} survives in the 2026 papers as the deployment model for inference; what the landmark set does not contain is any post-2022 work on the cloud's own abstractions---storage, consistency, multi-tenancy---of the standing of the Borg or Spanner papers.9495\subsection{Edge computing: a change of subject without a change of output}96\label{sec:p6-cloud-edge}9798Edge computing offers the sharpest contrast between what is cited and what is produced. Its baseline landmarks are the blockchain--IoT integration papers of 2018--2019, the \emph{Edge Intelligence} survey of \citet{zhou2019edgeintelligence}, and a cluster of COVID-era reviews of IoT applications that the classifier assigns to the category; the lineage runs from the definitions of \citet{shi2016edge} and \citet{satyanarayanan2017edge} through the edge-intelligence surveys \citep{deng2020edgeintelligence} to the 6G framing of edge AI \citep{letaief2022edgeai6g}. The post-period landmarks are surveys and visions about language models at the edge: mobile edge intelligence for LLMs, which positions the mobile network edge between on-device and cloud inference on grounds of privacy, latency and bandwidth \citep{qu2025meillm}; edge general intelligence with agentic AI, which proposes that distributed edge entities become autonomous agents with perception--reasoning--action loops \citep{zhang2026edgegeneral}; edge--cloud collaborative computing on distributed intelligence and model optimisation \citep{liu2026eccc}; and AI Flow, a device--edge--cloud framework for large models \citep{an2026aiflow}. The abstracts describe architectures, taxonomies and challenges; none reports a measured deployment, and the empirical question of where inference should sit on the device--edge--cloud continuum under energy, latency and privacy constraints is open.99100Against this change of subject, the category's time series barely moved. Edge computing has the smallest trend change at December 2022 of the eleven (\ItsSeriesCatEdgeComputingItsChatgptSlopeChange{} per month, $p = \ItsSeriesCatEdgeComputingItsChatgptSlopeP$), after a baseline growth of \CagrEdgeComputingCagrBaseline{} per year that had already raised its share from \CagrEdgeComputingShareTwoZeroOneEight{} to \CagrEdgeComputingShareTwoZeroTwoTwo{} by 2022; its post-period growth (\CagrEdgeComputingCagrPost) is lower than its baseline growth, its single selected break falls in 2025, and its references go to computer networks (\SysflowEdgeComputingPostToComputerNetworksShare\%) more than to itself (\SysflowEdgeComputingPostToEdgeComputingShare\%) in both periods. The LLM-at-the-edge literature is, on this evidence, a change in what the field writes surveys about, not a measurable change in its volume or its citation neighbourhood.101102\section{Computer networks}103\label{sec:p6-networks}104105Computer networks is the largest systems category and the one whose relative weight fell most, from \CagrComputerNetworksShareTwoZeroOneEight{} of primary output in 2018 to \CagrComputerNetworksShareTwoZeroTwoFive{} in 2025, with absolute output contracting at \CagrComputerNetworksCagrBaseline{} per year before 2022 and growing at \CagrComputerNetworksCagrPost{} after. It is also the category whose structural-change statistics are least compatible with a post-2022 event: its level change at December 2022 is positive (\ItsSeriesCatComputerNetworksItsChatgptLevelChange, $p = \ItsSeriesCatComputerNetworksItsChatgptLevelP$) but with a placebo $p$-value of \ItsSeriesCatComputerNetworksPlaceboChatgptLevelTPlaceboP, its trend change (\ItsSeriesCatComputerNetworksItsChatgptSlopeChange) is the smallest after edge computing, and its single selected Bai--Perron break precedes the event month (Table~\ref{tab:systems_categories}). The series was rising before ChatGPT.106107The most cited works of the period are overviews that set the agenda of the sixth generation of mobile systems: the 6G roadmap of \citet{wang2023road6g}, which surveys requirements, scenarios and testbeds while noting that the ITU-R vision was still under discussion, and the integrated sensing and communications overview of \citet{liu2022isac}, which treats the perceptive radio access network as a defining feature of 6G; both coincide with or precede the start of the window and continue the programme of \citet{dang2020sixg}. Machine learning had entered networking research well before the window, as the surveys of deep learning for IoT analytics and for mobile and wireless networking show \citep{zhang2019dlwireless}, and the programmable-network lineage of OpenFlow, P4 and in-network computation \citep{mckeown2008openflow,bosshart2014p4,kreutz2015sdn,hauser2023p4survey,sapio2021switchml} is cited throughout. What the 2026 landmarks add is a vocabulary: tutorials on large AI models and agentic AI for intelligent communications \citep{jiang2026lamagentic}, a review of energy, scalability, data and security in massive IoT \citep{cheikh2026massiveiot}, the edge-general-intelligence survey already cited, and a tutorial on pinching-antenna systems that belongs to the physical layer. The abstracts of the agentic-AI tutorials describe design principles and taxonomies; their empirical content cannot be assessed from abstracts, and they are provisional landmarks by the criterion of Section~\ref{sec:p6-scope}. The networks category also receives the references of edge computing and sustainable computing in larger proportion than those categories' own works, so that part of what the taxonomy records as networking output is infrastructure work that other communities cite. The established conclusion is modest: the field's agenda in 2023--2026 was 6G, and the arrival of large models is visible in its surveys rather than in its output or its citation structure.108109\section{Database systems and information retrieval}110\label{sec:p6-data}111112\subsection{Database systems: data for learning, and vector search}113\label{sec:p6-data-db}114115The database landmark list is the most contaminated by misclassification of the eleven, and what remains after exclusion is small but instructive. The baseline anchors are BLOCKBENCH, a data-processing view of private blockchains with a benchmarking framework \citep{dinh2018blockbench}, the documentation standard \emph{Datasheets for datasets} \citep{gebru2021datasheets}, and the GPU similarity-search paper that underlies the Faiss library \citep{johnson2021faiss}. The post-period landmarks are three: \emph{The Faiss Library} \citep{douze2026faiss}, whose abstract frames vector similarity search as the core functionality of vector databases and describes the library's indexing, compression and benchmarking principles; the data-centric AI survey of \citet{zha2025datacentric}, organised around training-data development, inference-data development and data maintenance; and an empirical study of the effect of six data-quality dimensions on nineteen machine-learning algorithms under polluted training and test data \citep{mohammed2025dataquality}. Database systems' trend change at December 2022 (\ItsSeriesCatDatabaseSystemsItsChatgptSlopeChange{} per month) and its single Bai--Perron break in late 2023 are unremarkable among the systems categories; its output contracted at \CagrDatabaseSystemsCagrBaseline{} per year before 2022 and grew at \CagrDatabaseSystemsCagrPost{} after, and its references go to machine-learning theory, natural language processing and computer vision in proportions that the two periods share.116117Two readings are consistent with this record and the corpus cannot separate them. The first is that the database community's most cited work of the period concerns the data that feed machine learning---its quality, documentation and maintenance---and the retrieval of embeddings, rather than transaction processing or query optimisation; this is what the metadata show. The second is that the engine-level work of the period is simply not cited enough within thirty-six months to reach a list ranked by short-window citations, which favours surveys. The bibliography records the community's own positions on both points: the vector-database surveys \citep{pan2024vectordb,taipalus2024vectordb} and the preprint of \citet{han2023vectordb}, the hierarchical navigable small-world index that most vector systems implement \citep{malkov2020hnsw}, the programme of machine learning for database internals \citep{li2021ai4db}, the assessment of large language models for data management by \citet{fernandez2023llmdata}, and the argument of \citet{stonebraker2024goesaround} that specialised vector stores will be absorbed by extended relational engines as earlier specialised systems were. Whether vector search belongs in dedicated systems or in extended relational engines is the contested question of the period, and it is untested at workload level: the landmark set contains library-level evidence and surveys, not comparisons of dedicated and extended systems under mixed structured and approximate-nearest-neighbour queries.118119\subsection{Information retrieval: from graph recommendation to retrieval augmentation}120\label{sec:p6-data-ir}121122Information retrieval shows the clearest change in what is cited of any systems category. The baseline landmarks are graph-neural-network recommendation and its surveys: neural graph collaborative filtering, knowledge-graph attention networks, graph neural networks for social recommendation, self-supervised graph learning, and LightGCN \citep{he2020lightgcn}, whose abstract reports that the feature transformation and non-linear activation inherited from graph classification contribute little to collaborative filtering and can degrade it; the surveys of \citet{wu2023gnnrec} and \citet{gao2023gnnrecsurvey} close the baseline. The post-period landmarks are retrieval-augmented generation and language models for recommendation: the survey of retrieval-augmented large language models of \citet{fan2024ragsurvey}, the \emph{ACM Computing Surveys} review of \citet{huang2026ragsurvey}, which organises the paradigm into pre-retrieval, retrieval, post-retrieval and generation stages, the survey of retrieval augmentation for generated content across modalities \citep{zhao2026ragaigc}, the surveys of recommender systems in the era of large language models \citep{zhao2024recsysllm,wu2024llmrec}, and C-Pack \citep{xiao2024cpack}, which releases a training corpus, a Chinese embedding benchmark and the BGE family of embedding models whose abstract reports state-of-the-art results on the English MTEB benchmark as well.123124The methodological lineage of this literature is continuous with 2020: dense passage retrieval \citep{karpukhin2020dpr}, retrieval-augmented generation itself \citep[preprint]{lewis2020rag}, late interaction in ColBERT \citep{khattab2020colbert}, and the probabilistic framework that these replaced \citep{robertson2009bm25}; the dense-retrieval survey of \citet{zhao2024denseretrieval} and the preprint survey of language models for information retrieval \citep{zhu2023llmir} document the transition, as does the preprint RAG survey of \citet{gao2023ragsurvey}. What is notable is that the landmark set contains no new retrieval architecture of 2023--2026 and consists, for the post period, of surveys and one resource. This may reflect the citation lag of recent model papers as much as the field's output, and the review records it as speculative. The corpus statistics are the strongest of the systems group: information retrieval has the largest trend change at December 2022 (\ItsSeriesCatInformationRetrievalItsChatgptSlopeChange{} per month), a negative level change (\ItsSeriesCatInformationRetrievalItsChatgptLevelChange, $p = \ItsSeriesCatInformationRetrievalItsChatgptLevelP$), two selected Bai--Perron breaks, output growth of \CagrInformationRetrievalCagrPost{} per year after a flat baseline (\CagrInformationRetrievalCagrBaseline), and it is the only systems category whose references to the large-language-model category become measurable, at \SysflowInformationRetrievalPostToLlmShare\% of post-period references against \SysflowInformationRetrievalBaselineToLlmShare\% before. Its self-share of references nevertheless remained at \SysflowInformationRetrievalPostToInformationRetrievalShare\%, and natural language processing remained its second destination (\SysflowInformationRetrievalPostToNlpShare\%). The placebo $p$-value for the trend (\ItsSeriesCatInformationRetrievalPlaceboChatgptSlopeTPlaceboP) forbids attributing the acceleration to the event date, and the open problem that the literature itself states is evaluative: no agreed protocol separates the contribution of retrieval quality from that of parametric memory and of generation in retrieval-augmented systems.125126\section{Quantum computing}127\label{sec:p6-quantum}128129Quantum computing is the systems category least connected to the others: \SysflowQuantumComputingPostToQuantumComputingShare\% of its post-period references go to works of its own category, applied cryptography being the next destination at \SysflowQuantumComputingPostToAppliedCryptographyShare\%. It is also the fastest-growing after 2022 (\CagrQuantumComputingCagrPost{} per year, after \CagrQuantumComputingCagrBaseline{} in the baseline), the only systems category whose share of computer-science output rose throughout (\CagrQuantumComputingShareTwoZeroOneEight{} to \CagrQuantumComputingShareTwoZeroTwoFive), and one whose level fell at December 2022 (\ItsSeriesCatQuantumComputingItsChatgptLevelChange, $p = \ItsSeriesCatQuantumComputingItsChatgptLevelP$). Its two selected Bai--Perron breaks fall at the beginning of 2022 and in the spring of 2025 (Table~\ref{tab:systems_categories}). These are the statistics of a field growing on its own momentum, and the landmark list explains why: the window contains a coherent experimental sequence that has no connection with language models.130131The baseline anchors are Preskill's framing of the noisy intermediate-scale era \citep{preskill2018nisq}, the supremacy experiment on a 53-qubit programmable superconducting processor \citep{arute2019supremacy}, the photonic advantage experiment, and the reviews of NISQ algorithms and variational methods \citep{bharti2022nisq,cerezo2021vqa}. The post-period landmarks are the error-correction milestones. \citet{googlequantum2023suppressing} report that a distance-5 surface-code logical qubit modestly outperforms an ensemble of distance-3 qubits, the first demonstration that adding physical qubits improves logical performance on that platform; \citet{bluvstein2024logical} report a programmable logical processor on reconfigurable neutral-atom arrays with up to 280 physical qubits, a zoned architecture and mid-circuit readout; and \citet{googlequantum2025belowthreshold} report two below-threshold surface-code memories on the Willow processor, a distance-7 code whose logical error rate is suppressed by a factor of $\Lambda = 2.14$ per increase of two in code distance, with a real-time decoder, and a logical memory beyond break-even. The figures are those of the abstracts and the articles are peer-reviewed. Two 2025 demonstrations extend the sequence to modular architectures: distributed quantum computation between two trapped-ion modules about two metres apart by deterministic gate teleportation over an optical link \citep{main2025distributedqc}, and Aurora, a 35-chip photonic machine networked over fibre that its authors describe as a sub-performant scale model demonstrating the integration of all primitive components \citep{aghaee2025aurora}. These are feasibility demonstrations; the rates and fidelities required for error correction across modules are not reported in the abstracts.132133The contested claim of the period is utility before fault tolerance. \citet{kim2023utility} report, on a 127-qubit processor, accurate expectation values for circuit volumes beyond brute-force classical computation and argue that this is evidence of utility in the pre-fault-tolerant era. The abstract restricts the claim to brute-force simulation. The classical-simulation responses that followed, which reproduced the reported quantities with tensor-network and related methods, are not in the landmark set and are not in the verified bibliography, and they are therefore not cited; the claim is recorded as contested, and the open problem is methodological---utility experiments need independently committed classical baselines fixed before publication, because the classical frontier moves after it. On the algorithmic side, the review of barren plateaus in variational quantum computing \citep{larocca2025barren}, for which the corpus holds no abstract, and quantum-selected configuration interaction \citep{kanno2026qsci}, which samples electron configurations on a noisy device and diagonalises the Hamiltonian classically in the selected subspace, represent the two responses to the trainability problem of variational algorithms: characterise the obstacle, or move the optimisation to the classical side. The software and vendor records---the Qiskit design paper \citep[preprint]{javadiabhari2024qiskit} and a retrospective of IBM's processors and industry applications \citep{abughanem2025ibmquantum}---complete a list in which, as the bibliography anticipates for Part~X, the post-quantum standards of 2024 \citep{nist2024fips203} and the lattice foundations \citep{regev2009lwe,shor1997factoring} are the only links to the rest of computer science.134135\section{Sustainable computing}136\label{sec:p6-sustain}137138Sustainable computing is small (\CagrSustainableComputingNTwoZeroTwoFive{} primary works in 2025), grew at \CagrSustainableComputingCagrPost{} per year after 2022 against \CagrSustainableComputingCagrBaseline{} before, has a positive level change at December 2022 (\ItsSeriesCatSustainableComputingItsChatgptLevelChange, $p = \ItsSeriesCatSustainableComputingItsChatgptLevelP$; placebo $p = \ItsSeriesCatSustainableComputingPlaceboChatgptLevelTPlaceboP$) and cites computer networks (\SysflowSustainableComputingPostToComputerNetworksShare\%) and cloud computing (\SysflowSustainableComputingPostToCloudComputingShare\%) more than or as much as itself (\SysflowSustainableComputingPostToSustainableComputingShare\%). Its baseline landmarks concern ICT for the sustainable development goals and data-centre traffic models; the lineage that the bibliography documents is the measurement of machine learning's footprint---the energy of training natural-language models \citep{strubell2019energy}, the green-AI argument \citep{schwartz2020greenai}, the carbon-tracking tools of \citet{henderson2020carbon}, the global data-centre energy accounting of \citet{masanet2020datacenters}---and the post-period landmarks continue it with a notable share of preprints and workshop papers.139140The measurement literature proceeds in three steps. Training: \citet{patterson2021carbon}, in a preprint, estimate the energy and CO$_2$-equivalent of T5, Meena, GShard, Switch Transformer and GPT-3 and report that sparsely activated models, data-centre location and hardware choice change emissions by large factors; \citet{luccioni2023bloom}, in a preprint, estimate BLOOM's training at 24.7 tonnes of CO$_2$-equivalent for dynamic consumption and 50.5 tonnes over the life cycle, and extend the accounting to inference through an API; \citet{dodge2022carbonintensity} argue that cloud providers should expose software carbon intensity to users. Inference and water: \citet{chien2023carbongenai} model generative-AI inference as a workload and compare request-direction policies for carbon; \citet{li2023thirsty}, first as a preprint and then in \emph{Communications of the ACM} \citep{li2025thirsty}, bring the water footprint of training and serving into the accounting, with projections of global withdrawal that are the authors' estimates; the reviews of green AI \citep{boloncanedo2024greenai} and green federated learning \citep{thakur2025greenfl} and the related argument of \citet{wu2022sustainableai}, a preprint, organise the mitigation side. Systems and policy: \citet{luccioni2024powerhungry} measure the energy of inference by task at the level of deployed models; \citet{luccioni2025jevons} argue that efficiency gains may raise total consumption through rebound effects and that life-cycle assessment must be joined to socio-economic analysis; \citet{devriesgao2026datacenters}, continuing \citet{devries2023energy}, document that operators' environmental disclosures do not separate AI from other workloads, so that AI's carbon and water footprints can only be approximated from aggregate data-centre metrics; \citet{chen2026aielectricity} review the electricity demand of AI data centres and its interaction with power-system planning, and \citet{kumar2026sustainabledc} review the infrastructures and energy-management options of sustainable data-centre operation; and TokenPowerBench \citep{niu2026tokenpowerbench} provides a benchmark that attributes inference energy to the prefill and decode phases of each request. The latter's abstract quotes industry reports to the effect that inference accounts for more than ninety per cent of total power; such figures recur across this literature and cannot be verified from the corpus. The field's own diagnosis is the open problem: fleet-scale footprints are modelled rather than measured because the data are not disclosed, and the rebound hypothesis has not been tested.141142\section{Synthesis}143\label{sec:p6-synthesis}144145Three statements are established by the corpus. The systems categories accelerated after 2022 as a group, and the acceleration is not specific to the ChatGPT date: every placebo test fails to single it out, and the break dates selected by the Bayesian information criterion are scattered across three years. The classical systems categories had been declining in absolute and relative terms since 2018 and recovered in absolute terms only; their share of computer science continued to fall, and reversion is a competing explanation of the recovery that this corpus cannot exclude. The citation flows among systems categories were stable between the two periods; the only measurable new destination is the references of information retrieval to the large-language-model category.146147Two statements are established by the metadata of the most cited works, and they must be read with the caveat that short citation windows favour surveys and that roughly one landmark record in five is misclassified. The most cited post-2022 works in computer architecture, operating systems, distributed computing and cloud computing are systems for training and serving large language models---graph compilation, phase splitting, key--value-cache management, GPU sharing, consumer-GPU inference, inference scheduling---and the most cited works in information retrieval, edge computing and networking are surveys and tutorials about language models. Quantum computing and sustainable computing follow their own trajectories, the former an experimental programme in error correction and modular architectures, the latter a measurement literature on the footprint of machine learning in which several foundational items are preprints.148149What cannot be established from this evidence is also to be stated. Whether research on language-model infrastructure explains the volume recovery of the classical categories is unknown, because the share of that research in each category's output is not measured here; a topic-level decomposition is the test. Whether any of the systems designs reviewed---paged or contiguous caches, phase-split or co-located serving, hot--cold partitioning---is superior to its competitors is unknown, because each is evaluated by its own authors on its own hardware. Whether photonic and neuromorphic efficiency figures hold at system level is unknown, because no abstract reports end-to-end energy under a shared benchmark. Whether pre-fault-tolerant quantum processors deliver utility is contested and moves with the classical frontier. These, together with the evaluation of retrieval augmentation, the placement of inference on the device--edge--cloud continuum, the disclosure of AI's energy and water use, and the validation of the taxonomy's systems categories against venue streams, are carried forward to Part~X.150