Technology Companies — A Ranking by What Binds Them
Technology Companies — A Ranking by What Binds Them
Frontiers article in the Harmonism cascade. The company-level ranking beside The Sovereign Stack, which judges projects, and Open-Weight AI Models Ranked — Censorship, Capability and Bias, which judges models. Figures as of October 8, 2026. See also: The Telos of Technology, The Techno-Dharmic Manifesto, Open Source and Harmonism, Finance and Wealth.
In April 2026 Meta moved its frontier work to a closed model, Muse Spark, after three years of releasing its Llama models as open weights. Nothing in Meta’s conduct before April announced the turn, and nothing in Meta’s structure forbade it. A company’s present behaviour records what its owners currently find profitable. What a practitioner can rely on is whatever binds the behaviour that comes next.
A technology company serves the sovereignty The Sovereign Substrate describes only as far as two things hold together: a structure binds its future conduct, and its product leaves the buyer owning rather than renting. Three structures do the binding. A release under an open licence cannot be withdrawn once it is made; ownership held by a foundation or a purpose trust cannot be sold to whoever bids highest; and revenue from things sold outright pays a company to keep selling them. Ranked by those criteria, the field inverts its market order. The companies a practitioner can trust furthest are small, nonprofit, foundation-held or publicly funded, and almost none of them can be bought on a stock exchange.
1. What Binds a Company, and Toward What
Two axes order the ranking, and each fails without the other.
The first axis is binding. A licence binds the artefact: when Google released Gemma 4 under Apache 2.0 on April 2, 2026, dropping the custom licence whose prohibited-use policy it could rewrite at will, those weights became permanently usable on Apache terms whatever Google decides about Gemma 5. Ownership binds the company. Ernst Abbe placed Carl Zeiss in a foundation in 1889 so that its profits would serve research and the workforce rather than shareholders, and Bosch has been held the same way for decades; in June 2024 Proton’s founders gave their shares to the Proton Foundation, a Swiss nonprofit with no shareholders, and no change of control at Proton can now happen without its consent. Revenue binds incentive. A company paid when the customer buys a machine profits from the machine working in the customer’s house, while a company paid per query profits from the customer never being able to leave.
Revenue predicts the other two more reliably than either predicts it. Mozilla is foundation-owned and its browser is open source, and most of its revenue has come from Google paying to be Firefox’s default search engine, so the organisation’s survival has rested on the company whose model of the web Firefox exists to resist.
The second axis is direction, and binding alone does not supply it. The Novo Nordisk Foundation holds the controlling stake in Novo Nordisk under a structure as durable as any in Europe, and the structure binds the company firmly to the pharmaceutical paradigm Big Pharma diagnoses. A locked structure keeps a company on its course. Whether the course leads anywhere worth going is the question The Telos of Technology answers: ownership, not subscription, and technology placed under Stewardship because it must serve Dharma.
Harmonia holds Lodestone to the same two axes. Its desktop app runs Lodestone on the practitioner’s own hardware, the one place no one else reads the conversation; its hosted companion rents open-weight models and says so, because rented inference, attested or not, rests on a promise (Inference Sovereignty); the direction is set out in Harmonia’s AI Infrastructure. The rule binds by architecture rather than by promise, because a promise is the one thing an operator can revise without anyone noticing.
The grades follow The Sovereign Stack. Aligned means a binding structure and an owning direction together. Adjacent means one of the two, or both in part. Not aligned means neither, though a non-aligned company’s hardware or software can still be usable with its closed layer understood.
2. Why Present Conduct Does Not Count
Most structural changes of the last three years in this field ran one way, from open toward closed or from bound toward sellable, with two exceptions running the other way. Red Hat, by then owned by IBM, closed public access to the RHEL source in June 2023. HashiCorp moved Terraform to the Business Source License that August. Redis left its open licence in March 2024 and came back to AGPLv3 in 2025 after distributions dropped it, and Google moved Gemma to Apache 2.0 in April 2026. On October 28, 2025 OpenAI completed the recapitalisation that turned the nonprofit created in 2015 to control it into the OpenAI Foundation, holding about 26% of OpenAI Group PBC beside Microsoft’s roughly 27%. Meta closed its frontier in April 2026. On February 2, 2026 SpaceX announced that it would acquire xAI, and X with it, in an all-stock deal valuing xAI at $250 billion. In June 2026 Qualcomm was reported in advanced talks to buy Tenstorrent.
What survived each change was the artefact already released under a licence that could not be recalled. The RHEL rebuilds found their way through CentOS Stream. OpenTofu forked Terraform at its last open version. Every Llama checkpoint published before April 2026 still downloads and runs. The engine in llama.cpp stayed MIT-licensed when its maintainers joined Hugging Face in February 2026.
The corpus carries the lesson in its own records. Inference Sovereignty names Llama among the Western open-weight lines and records in the same entry that Meta moved its frontier to a closed model, so a single paragraph of the tier map now holds both the grade and its expiry. A grade given for conduct expires with the conduct. A grade given for structure lasts as long as the structure, and a grade given for a released artefact lasts as long as someone keeps a copy.
3. Silicon Makers
Open hardware and capable hardware are, in 2026, sold by different companies, and the ordering below follows binding before capability.
MNT Research and Pine64 are aligned. MNT publishes the schematics, firmware, mainboard and mechanical drawings of the Reform laptop, and Pine64 sells hackable phones and laptops that run mainline Linux; a published schematic cannot be unpublished. Neither makes a machine that holds a hundred-billion-parameter model, which is the price of the grade at this layer.
Tenstorrent is adjacent and stands closest to alignment among makers of AI accelerators. Its Blackhole cards sell to individuals, the p150a with 32 gigabytes for $1,399 and the p100a with 28 for $999, and its software is published under Apache 2.0 from the low-level programming framework up through the inference server. Its ownership binds nothing: it is venture-funded, and the reported Qualcomm talks would put its roadmap in the hands of a company whose revenue comes mostly from phone chips. What it has released stays open whoever buys it.
Purism, Framework and System76 are adjacent, each in a different way. Purism reincorporated as a social purpose corporation in 2017, a legal form that writes its mission into its charter, and ships coreboot and the tamper-evident Heads firmware on its Librem laptops. Framework builds laptops and a desktop meant to be opened, sells every part and publishes the repair guides; its 128-gigabyte Framework Desktop solders the memory because AMD’s Strix Halo bus cannot run through a socket. System76 ships Linux desktops with open firmware on selected models and room for whatever accelerator the owner installs. All three sell outright.
NovaCustom, Star Labs and Nitrokey are adjacent and stand closest to alignment among makers whose machines run a 27-billion-parameter model. Their binding lives in the firmware rather than the ownership: NovaCustom in the Netherlands and Nitrokey in Germany ship Dasharo, the coreboot-based firmware the Polish firm 3mdeb maintains, NovaCustom with Heads as an option and Intel’s Management Engine disabled on request, and Star Labs in Britain ships coreboot on its StarBook laptops. All three sell outright. The binary blobs Intel and AMD require remain, and none of the three yet sells a machine with the memory for the hundred-billion class; coreboot reached Strix Halo on June 25, 2026, but only on AMD’s reference board.
AMD is adjacent. It publishes most of ROCm, its compute stack, under MIT and Apache terms, keeps a few profiling and compiler components as proprietary binaries, and ships its graphics drivers inside the Linux kernel. Its Strix Halo chip is the cheapest route to 128 gigabytes of memory a model can use. Intel is adjacent on the same grounds, with open Linux drivers and the cheapest graphics memory per dollar on its Arc Pro B60.
Apple’s hardware is adjacent and the company is not aligned. Its incentive points toward local inference more than any other large maker’s, since it sells machines outright and has staked its own AI on running inside them, and the M5 Ultra Mac Studio is the single box that holds the largest open models. Everything else about the company points the other way: the operating system is closed, the memory sits in the chip package, storage swapped by its owner fails without Apple’s restore, and Linux reaches Apple Silicon only through the volunteer Asahi project, which has released nothing yet for M4 or M5. Lodestone’s desktop app is notarized through Apple’s Developer Program at $99 a year, because a Mac otherwise warns its own owner against opening a program Apple has not checked.
Lenovo, the Taiwanese brands and the Chinese mini-PC makers are not aligned and usable as machines. Lenovo’s ThinkPads are well documented and certified for Linux, and in 2015 Lenovo shipped both the Superfish adware and a Service Engine that reinstalled its software from the firmware after the disk was wiped. ASUS, GMKtec and Minisforum sell the same Strix Halo chip with 128 gigabytes as Framework, without Framework’s repair commitment or published guides. Huawei is not aligned and not usable outside China for this work: its PCs are moving to HarmonyOS, a closed system on its own kernel, and its Ascend accelerators go to data centres, though its August 2025 decision to open-source CANN, its alternative to CUDA, would make its stack more readable than NVIDIA’s if it ships in full. Country of origin decides less than the firmware does. Nearly every chip on this list is made by TSMC in Taiwan with Korean memory, and firmware no outsider can read, Intel’s Management Engine and AMD’s Platform Security Processor among it, is a risk whichever flag the company flies.
NVIDIA is not aligned as a maker and usable as a machine. The RTX 5090 and the DGX Spark are the fastest hardware a practitioner can own, CUDA is proprietary, and about nine dollars in ten of NVIDIA’s revenue in the quarter to January 2026 came from the data centres local inference routes around.
4. Model Makers
At the model layer the binding structure is the licence of the release, and the strongest release is the one that publishes the training data and code alongside the weights, because only then can a practitioner read the hand the model carries.
The Allen Institute for AI is aligned. It is a nonprofit, and OLMo 3 is fully open: weights, the Dolma 3 corpus, the training code and the intermediate checkpoints. Lodestone in the browser already ships the earlier OLMo 2 7B. The Swiss AI Initiative is aligned on the same grounds. ETH Zurich, EPFL and the Swiss National Supercomputing Centre released Apertus on September 2, 2025 under Apache 2.0 with its training data and recipes open, and Proton’s Lumo now serves Apertus 1.5. Public money built it, and no investor can sell it.
Mistral, DeepSeek and Alibaba are adjacent. Mistral releases Mistral Small 3.2 and Mistral Large 3 under Apache 2.0, keeps other models closed and answers to venture investors. DeepSeek releases its V3 and V4 weights under MIT and is owned by the hedge fund High-Flyer; Alibaba releases most of Qwen under Apache 2.0 and is listed in Hong Kong and New York. Both Chinese labs carry the refusals Open-Weight AI Models Ranked — Censorship, Capability and Bias measures around topics sensitive to the Chinese state, and Perplexity needed about forty thousand prompts of new training to take them out of DeepSeek R1. Google’s Gemma 4 belongs with them as an artefact; Google does not, since its frontier stays closed and its revenue comes from the attention economy.
OpenAI, Meta, Anthropic and xAI are not aligned. OpenAI released gpt-oss under Apache 2.0 and rents everything that matters through its API, and its recapitalisation demonstrated how far a founding structure can be renegotiated. Meta’s closed turn needs no further comment. Anthropic carries the strongest governance structure of the four, a public benefit corporation with a Long-Term Benefit Trust electing a growing share of its board, but its stockholders can amend the Trust by supermajority, and its product is access rented by the token to weights nobody outside can inspect. A companion built on that rented access stops the day the account does, whatever its practitioners need, and Lodestone’s own production companion ran on Claude until its migration in October 2026. xAI’s hand stands apart from the other Western labs’, though its weights stay closed, in Tier B of Inference Sovereignty, and the company itself now belongs to SpaceX.
5. Software and Service Companies
At the service layer even the best structure cannot fix one architectural fact: a service run by a single operator can be changed by that operator, so the grades here top out at adjacent unless the software also runs on the practitioner’s own machine.
Signal is adjacent, and the best-governed service in this section. The Signal Technology Foundation is a nonprofit funded by donations, the protocol and clients are open source and repeatedly audited, and the weak points are the phone number registration still requires and the servers only Signal runs. Proton is adjacent. Its foundation now controls it, its clients are open source and its jurisdiction is Swiss. Its Lumo assistant runs open models on servers Proton controls under a no-logs policy that the architecture cannot yet prove, which places it in the promissory regime Inference Sovereignty distinguishes from attested inference. Mozilla is adjacent with the warning already given: the structure is sound, and the money comes from the adversary.
Hugging Face is adjacent and has become the field’s single point of concentration. The registry from which nearly every open model is downloaded and, since February 2026, the team that maintains llama.cpp belong to one venture-funded company. Lodestone’s desktop app checks every model file it downloads against the hash Hugging Face publishes, which means Harmonia’s own sovereign path currently runs through that company’s servers. The engine’s MIT licence keeps the concentration survivable on one condition: that practitioners and institutions keep their own copies of the weights they rely on.
Ollama and Venice are adjacent. Ollama’s code is MIT-licensed and its company is venture-backed; Venice serves open models, runs a subset inside attested enclaves and remains a single private operator. LM Studio is not aligned and usable: local-first in posture, proprietary in code. Jan is the open alternative.
6. How to Use the Ranking
Buy artefacts rather than promises. A machine that opens, a model whose licence was granted the day it was released and a protocol anyone can run keep their properties after the company that made them changes owner, strategy or mind. Prefer the company whose revenue arrives when you buy, and treat every grade above as a judgement on what the company will do next rather than a reward for what it has done.
Keep copies. A release binds only as long as someone holds it: the weights on your own disk, the inference engine at a version you have built, the installer you verified. The Practitioner’s Disciplines call this backing up what cannot be lost, and at the model layer it is also what keeps the commons from depending on the one company that hosts it.
Hold shares, if you hold them, as what they are. Finance and Wealth § The Capture Gradient places a listed share among the captured forms, title without stewardship, ranked higher when held directly and higher again when the company’s ownership binds its conduct. The ranking above adds a fact the gradient anticipates: Ai2, the Swiss AI Initiative, Signal, Proton, MNT, NovaCustom, Star Labs, Framework, Purism and Tenstorrent cannot be bought on an exchange, and the listed companies here (Apple, AMD, Intel, Lenovo, Alibaba, Google, Meta, NVIDIA) are adjacent at best. The New Acre gives the deeper reason the aligned names are scarce on exchanges: if production moves into households, its value arrives as rent no longer paid, which accrues to the owner of the machine and to no company’s shareholders. None of this is a recommendation to buy or sell anything.
Four kinds of claim sit in the ranking. Harmonism holds that stewardship, not title, is the right relation to a productive asset and that technology must be owned rather than rented to serve Dharma. The evidence of the last three years supports the narrower claim that conduct is revocable and released artefacts are not. The traditions witness the same pattern in the endowment, the waqf and the monastic estate, each of which made dissolution a loss rather than a payout. Whether a foundation-held or publicly funded maker can raise the capital the frontier now demands, or whether the open frontier will always run a few months behind the closed one, remains open.
The ranking will be re-run each quarter, and whenever a company on it changes owner. A change of owner is re-graded on the new structure, never on the press release that announced it.
The companies that build the tools of the coming century are deciding now whether those tools will be held by the households that use them or rented from the companies that make them. A practitioner cannot decide that for them. A practitioner can decide where to put money, attention and trust, and can put them where a structure, and not a mood, holds the company to the side of ownership.
See also: The Sovereign Stack, Open-Weight AI Models Ranked — Censorship, Capability and Bias, Inference Sovereignty, Lodestone, Open Source and Harmonism, The Telos of Technology, The Techno-Dharmic Manifesto, Finance and Wealth, Technology and Tools, The New Acre.