Compare moats
Up to three covered companies, band by band. Every call is a curated editorial judgment, never a disclosed figure — and every band carries its cited basis.
| MediaTek | Cipher Mining | MiniMax | |
|---|---|---|---|
| Moat rating | narrow Counterpoint's public quarterly AP-SoC table, published 8 September 2026, shows MediaTek first in all ten quarters it covers (Q1 2024 through Q2 2026), never below 31%, with the nearest rival never above 30% -- a sustained position, not one good quarter. What the same table does not show is a widening one: the share oscillates inside a 31% to 41% band, the Q1 readings step down 41% (2024), 38% (2025), 33% (2026), and the 31% of Q2 2026 matches the lows already set in Q4 2024 and Q4 2025 rather than breaking them; Q2-on-Q2 the series runs 34%, 36%, 31%. Counterpoint also records MediaTek's shipments down more than 30% year-over-year in the quarter, the mainstream and entry tiers hit by the memory shortage and the premium tier by a 'slowdown in sales of Dimensity 9000 series powered devices.' MediaTek's own 2Q26 remarks show the gross line holding and the operating line not: gross margin 46.2%, down only 0.1 percentage points sequentially and guided at 46% plus or minus 1.5 for Q3, while operating income fell 22.2% year-over-year on revenue up 1.2% as operating expenses rose to NT$47.4 billion from NT$44.5 billion. A first place held for ten quarters without extending it, and earning less on it, is defensible rather than compounding. | none Cipher (renamed Cipher Digital Inc. on 20 February 2026) holds signed, credit-supported leases but has not yet shown that they amount to a durable advantage. Its 2025 Form 10-K (filed 2026-02-24) says "Through the end of 2025, our revenue has been derived from mining bitcoin". Its Q2 2026 business update (2026-08-04, Exhibit 99.1, https://www.sec.gov/Archives/edgar/data/1819989/000181998926000038/q226_earningsxprxdraftxvf.htm) reported "Q2 2026 Revenue of $25 million", and its first HPC rent began only in August 2026 at Black Pearl. The 10-K's own risk factors describe a contested market. They say "There has been an increasing number of businesses constructing HPC data centers, which has resulted in increasing competition and pricing pressure that may cause us to reduce our pricing in order to remain competitive", that "our data centers are single-tenant properties", and that lessees "will have the right to terminate the lease if there are significant delays in the completion of construction". At Barber Lake, where the 10-K targeted Phase I delivery "by September 30, 2026", a September 2026 amendment, made "In connection with change orders and the continued evolution of tenant requirements", moved data-hall deliveries to the fourth quarter of 2026 through the first quarter of 2027. Under it, Cipher "will bear the first $359.3 million of costs in excess of the initial budgeted amount" (2026-09-25). There is real counter-evidence: a 15-year Amazon lease, a Google-backstopped Fluidstack lease whose contracted life a "leading AI lab" extended to 20 years, and a third lease with an investment-grade hyperscale tenant. Until that capacity is delivered and paying rent, though, a moat is not shown, so the band is none. | none MiniMax is growing fast at a heavy loss, with nothing yet shown to protect its position. Its interim results announcement for the six months to 2026-06-30 reports revenue of US$116.6 million, up 283.1%, against a loss for the period of US$358.0 million and research and development expenses of US$296.9 million, with gross profit margin at 17.9%. Its prospectus (https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1231/2025123100025.pdf) warns that competition may make it "more difficult for us to differentiate our products, maintain pricing power, and achieve sustainable profitability". Independent trackers show its standing has to be won again with each release. The prospectus says MiniMax-M2 became "a top three foundation model worldwide by daily token usage on OpenRouter" within the first week of its October 2025 launch, yet OpenRouter's rankings (https://openrouter.ai/rankings, as read 2026-10-10, usage data through Oct 9, 2026) list no MiniMax model among the top 20 by weekly usage. Artificial Analysis's leaderboard (https://artificialanalysis.ai/leaderboards/models, read 2026-10-10) scores MiniMax-M3 at 29 on its Intelligence Index, against 45 for Z AI's GLM-5.3 (max) and 44 for Kimi K3 (max). A loss-making developer whose usage rank resets with each model cycle has no moat yet. |
| Moat type | intangibles ip MediaTek manufactures nothing itself -- its 2Q26 remarks describe 'deep design-technology co-optimization (DTCO) with TSMC' and close collaboration with advanced-packaging partners -- so the durable asset cannot be plant scale, and the remarks do not claim customer lock-in either. They locate it in a reusable design library: 'Building on our industry-leading IP portfolio, we offer best-in-class, pre-validated subsystems solutions for memory, I/O, and connectivity' -- which, the company says, 'largely reduces design complexity and shortens time to market for our data center customers'. The same remarks itemise 448G SerDes with a Co-Package Copper system solution, CPO development on TSMC's COUPE platform, and an end-to-end 3.5D platform with 'best-in-class 3.5D IP, packaging, and design flows,' and say design wins come 'backed by our industry-leading IPs, deep ecosystem partnerships, and proven execution capabilities'. The spending behind that library shows in the cost line: operating expenses of NT$47.4 billion in the quarter against NT$44.5 billion a year earlier, with the Q3 operating expense ratio guided at 31% plus or minus 2 points. It is the same library MediaTek re-amortises across mobile, edge and now data-center ASIC. | none No moat source is demonstrated yet. Switching costs are the most likely candidate, since the leases are long (15 years at Black Pearl; Barber Lake now a 20-year contracted life) and each site has a single tenant. But the 10-K says tenant guarantees "will only be effective after rent commencement under such leases and are subject to certain limitations", and by August 2026 rent had begun at only one site. The advantages the 10-K claims are the company's own. It speaks of "industry-leading expertise in originating and securing industrial-scale, greenfield data center sites" and of securing West Texas land "on more favorable terms than in more established data center markets". The same document says "there is significant competition for power capacity and energized facilities". Intellectual property is modest: "four granted United States patents and one issued patent in Taiwan". With 66 full-time employees, Cipher has no scale advantage over the competitors it names: CoreWeave, Digital Realty, Equinix, Vantage Data Centers and Aligned Data Centers. | none None of the usual sources of a moat is established in the record. Much of the technology is published: the prospectus says MiniMax makes "certain of our models and products available on an open-source basis", which allows "third parties, including competitors, to access, use, modify, or redistribute them, which could limit our ability to commercialize those technologies or differentiate ourselves in the marketplace", and the interim results say MiniMax H3 was released "with open weights". On Artificial Analysis's video leaderboard (https://artificialanalysis.ai/video/leaderboard/text-to-video, read 2026-10-10), the entry ranked just below MiniMax's own H3 is "MiniMax H3 Max", listed under the creator Fal and marked "Based on MiniMax H3". On the API side, KrASIA's report on IDC data (https://kr-asia.com/how-bytedances-volcano-engine-holds-nearly-half-of-chinas-maas-market, 2026-05-13) notes that "In theory, developers only needed to change a few lines of code to replace the underlying model or switch cloud platforms", though it reports that the market leader, ByteDance's Volcano Engine, held its share in 2025, at 49.5%; it does not name MiniMax. Scale belongs to others: CIC, the industry consultant MiniMax commissioned, ranked it tenth among global foundation model companies by 2024 model-based revenue, "with a market share of 0.3%", and the prospectus concedes that many competitors have "greater access to data, talent, and computing infrastructure". |
| Leadership | co leader Counterpoint has MediaTek ahead of every rival in all ten quarters of its public table -- 31% against Qualcomm's 23% and Apple's 19% in Q2 2026 -- but that count covers smartphone application processors only, and MediaTek's 2Q26 remarks put mobile phone at 41% of group revenue, no longer its largest business. Smart Edge Platforms is larger at 53%, and while the remarks claim 'continued market share gains across connectivity, computing and automotive products', no independent count of those positions appears in the sources used here. In data center MediaTek holds nothing yet: production of its first AI accelerator ASIC is scheduled for the fourth quarter of 2026, and 15% to 20% is a 2027 target raised from 10% to 15% a quarter earlier. Inside its core market the flagship tier is a stated plan rather than a claimed position -- a 2nm SoC launching in the third quarter and, 'together with ongoing engagements with global players, we will increase our presence continuously in the flagship smartphone market in 2027' -- while Counterpoint records MediaTek's premium-tier shipments declining on a 'slowdown in sales of Dimensity 9000 series powered devices.' | fast follower No independent share or rank was found. The 10-K names CoreWeave, Digital Realty, Equinix, Vantage Data Centers and Aligned Data Centers as competitors, along with miners that have "signed leases with hyperscalers and HPC tenants": Hut 8, IREN, TeraWulf, Core Scientific and Applied Digital. Cipher has signed hyperscale-grade tenants (Amazon; Fluidstack with a Google backstop; a third investment-grade hyperscaler) but delivered its first HPC capacity only in August 2026 (Q2 2026 update). That makes it a fast follower. Its self-description as "a leading developer, owner, and operator of industrial-scale data centers" is the company's own claim and is not counted. | behind Independent rankings put MiniMax behind the leading Chinese labs in language and speech, and near the front only in video. Artificial Analysis's text-to-video leaderboard (https://artificialanalysis.ai/video/leaderboard/text-to-video, read 2026-10-10) shows MiniMax H3 (768p) fourth with an Elo of 1137, against 1156 for Alibaba's Wan 3.0. Its language-model leaderboard (https://artificialanalysis.ai/leaderboards/models, read 2026-10-10) scores MiniMax-M3 at 29 on the Intelligence Index, against 46 for Xiaomi's MiMo-V2.6-Pro, 45 for Z AI's GLM-5.3 (max) and Alibaba's Qwen3.8 Max (0902), 44 for Kimi K3 (max) and 39 for DeepSeek V4.1 Flash (max), and its text-to-speech leaderboard, read the same day, ranks Speech 2.8 HD 19th. OpenRouter's top 20 models by weekly usage, as read 2026-10-10 (usage data through Oct 9, 2026; https://openrouter.ai/rankings), include models from DeepSeek, Z.ai, Xiaomi, Tencent, Moonshot AI and StepFun, but none from MiniMax. The only revenue ranking is commissioned: CIC put MiniMax tenth among foundation model companies globally by 2024 model-based revenue, "with a market share of 0.3%", and fourth among pureplay foundation model companies. One near-front video model does not offset a language model scoring 29 against 44 to 46 for the best Chinese rivals, no top-20 usage, a 19th-ranked speech model and a 0.3% revenue share: MiniMax is behind the leaders, not at parity with them. |
| Pricing power | moderate The 2Q26 remarks describe price as cost pass-through rather than capture: 'As rising costs across the supply chain have become an industry-wide reality, we are taking pricing actions to ensure these increases are appropriately reflected in our product pricing', with the full-year aim, 'through our disciplined pricing strategy', to 'deliver our full year gross margin within the current quarter guidance range' -- holding a band, not widening it. The realised figures show that defence working: gross margin 46.2%, down 0.1 percentage points sequentially, with Q3 guided at 46% plus or minus 1.5 points. The 2.9-point year-over-year decline is not an erosion rate; the company states it 'was mainly due to a one-time benefit in the year-ago quarter.' Holding a mid-40s gross margin through a market the company expects to decline 'by about 15% in units this year' is genuine discipline, but passing rising costs through to defend a guided band is not the ability to price above it. | weak The 10-K says competition "has resulted in increasing competition and pricing pressure that may cause us to reduce our pricing in order to remain competitive". It adds that if customers cut usage "we may be compelled to lower our prices or risk losing a significant customer". At Barber Lake, Cipher agreed to absorb "the first $359.3 million of costs in excess of the initial budgeted amount", with the tenant reimbursing "50% of any such costs above that amount" (2026-09-25, Exhibit 99.1, https://www.sec.gov/Archives/edgar/data/1819989/000181998926000043/barberlakeleaseamendmentpr.htm). Its one input-cost edge is specific to mining: Odessa's power costs about 2.8 c/kWh under the Luminant contract, available "until at least July 2027". A landlord that concedes cost overruns to its tenants has weak pricing power. | weak MiniMax competes on price and its margin is thin. Gross profit margin was 17.9% in the six months to 2026-06-30, up from 12.1%, which the interim results put down to "improving infrastructure efficiency". Its annual results announcement (https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0302/2026030202837.pdf) sells M2.5 on cost, saying that running it "continuously for one hour at an output speed of 100 tokens per second costs only one U.S. dollar". Artificial Analysis's video leaderboard, as read 2026-10-10, lists API pricing of $4.80 per minute for MiniMax H3 (768p), against $12.00 for Wan 3.0 and $34.12 for Dreamina Seedance 2.5. The prospectus says "our models' competitiveness, is directly related to our models' market pricing and demand" and warns that "we may be required to reduce our prices or offer alternative pricing models", and the July 2026 placing announcement says "prices of high-performance computing hardware have continued to rise". |
| Summary | MediaTek is the largest supplier of smartphone application processors by units and has been for as long as Counterpoint's public quarterly table runs. Across all ten quarters it covers, Q1 2024 through Q2 2026, MediaTek is first every time and never below 31%, while the nearest rival never clears 30%; in Q2 2026 it holds 31% against Qualcomm's 23%, Apple's 19%, UNISOC's 13%, Samsung's 9% and HiSilicon's 5%. That is a real, independently counted position, and it is the only part of MediaTek's business for which such a count exists in the sources used here. The same table is the clearest evidence that the position is being defended rather than extended. MediaTek's share moves inside a 31% to 41% band across the ten quarters, and the top of that band has come down -- 41% in Q1 2024, 38% in Q1 2025, 33% in Q1 2026 -- while the 31% of Q2 2026 matches the lows already set in Q4 2024 and Q4 2025 rather than breaking them; measured like for like against the same quarter a year earlier the series runs 34%, 36%, 31%. Counterpoint records MediaTek's shipments falling more than 30% year-over-year in the quarter -- hit in the mainstream and entry tiers by the memory shortage, and in the premium tier by a slowdown in Dimensity 9000 series devices -- while Samsung's Exynos climbed to 9% on Galaxy S26 base variants and mid-tier A-series wins. MediaTek's own 2Q26 remarks show the gross line holding and the operating line not: gross margin of 46.2%, down 0.1 percentage points sequentially and guided at 46% plus or minus 1.5 for the third quarter, against operating income down 22.2% year-over-year on revenue up 1.2%, with operating expenses of NT$47.4 billion against NT$44.5 billion a year earlier. Unit leadership held while the return on it did not. What MediaTek re-uses is design IP. It manufactures nothing itself -- the remarks describe deep design-technology co-optimization with TSMC and close work with advanced-packaging partners -- and they locate the durable asset in a pre-validated library of memory, I/O and connectivity subsystems, 448G SerDes with a Co-Package Copper system solution, CPO development on TSMC's COUPE platform, and an end-to-end 3.5D platform. The spending behind that library is visible in the cost line: operating expenses of NT$47.4 billion in the quarter against NT$44.5 billion a year earlier, with the Q3 operating expense ratio guided at 31% plus or minus 2 points against a gross margin guided at 46% plus or minus 1.5 -- an operating cost base running near a third of revenue. That library is now pointed at two adjacencies. Smart Edge Platforms -- connectivity, computing, automotive and TV SoCs -- is now the larger business, at 53% of revenue in Q2 against mobile's 41%, and grew 26% year-over-year while mobile fell 20%. And the company says it has built its first AI accelerator ASIC for a major US cloud customer, with production scheduled for the fourth quarter of 2026. One adjacency is a target and one is a contribution, and this profile treats them as exactly that. The data center share figure of 15% to 20% is a 2027 target against a served market MediaTek itself sizes at US$80 billion, raised from 10% to 15% only one quarter earlier, and production of the first ASIC was still scheduled for the fourth quarter at the date of the remarks. The RTX Spark AI PC work is corroborated by NVIDIA's own announcement, which credits MediaTek with collaborating on the custom CPU design and contributing to its power efficiency, performance and connectivity -- but the superchip, its Grace CPU, its NVLink-C2C interconnect and its software stack are NVIDIA's, and the systems ship from PC makers, so what MediaTek holds is a design contribution rather than a franchise with a share to count. The claimed share gains in connectivity, computing and automotive carry no number and no third-party count. The honest reading is a company with one durable, independently verified franchise it is holding rather than extending, funding a credible but still unproven move into higher-value silicon on the strength of its IP portfolio. | Cipher built bitcoin mining data centres in Texas and is now developing single-tenant AI and HPC campuses for lease to hyperscalers. Its 10-K reports a portfolio of "4.2 gigawatts ("GW") of capacity across 10 sites". It lists a 15-year Amazon Web Services lease for about 300 MW of turnkey capacity at Black Pearl and a Fluidstack lease at Barber Lake (300 MW gross) under which Google "has agreed to backstop certain obligations of Fluidstack". It describes bitcoin mining at Odessa on a Luminant power contract at about 2.8 c/kWh. In 2026 it signed a third campus lease "with an investment-grade Hyperscale tenant" (Q1 2026 update, 2026-05-05), delivered first Black Pearl capacity in August "two months ahead of the original schedule" with rent commenced, and fully funded its Stingray development with a bond (Q2 2026 update). On 2026-09-25 it said Barber Lake's contracted life was extended from 10 to 20 years, taking contracted revenue at the site "from $3.8 billion to over $9 billion". The 10-K says Odessa was "the first bitcoin mining data center awarded the Management and Operations, or M&O, Stamp of Approval award from the Uptime Institute", which is independent recognition of how the company operates. Against this, the latest quarter's revenue was still bitcoin mining and fell to $25 million. The 10-K describes growing "competition and pricing pressure", single-tenant concentration, and termination rights for construction delays. The Barber Lake schedule was reset, and Cipher absorbs the first $359.3 million of cost overruns. Cipher's contracted pipeline is substantial, but a competitive advantage is not yet demonstrated, so it is rated as having no moat. That could change to narrow once its leased campuses are delivered and paying rent. | MiniMax develops models across language, video, speech and music and sells them through its Open Platform and its own consumer apps, including MiniMax Agent, Hailuo AI, Talkie and Xingye; it listed in Hong Kong in January 2026 as a Specialist Technology Company under Chapter 18C of the Listing Rules. Growth has been steep. Per its interim results announcement, revenue in the six months to 2026-06-30 was US$116.6 million, more than its US$79.0 million for all of 2025; Open Platform and other AI-based enterprise services revenue rose 703.1% to US$73.9 million, 63.4% of the total, on API call volumes and "the rapid adoption of our Token Plan"; and 60.8% of revenue came from outside Chinese mainland. Independent trackers show a mixed technical position. On Artificial Analysis's text-to-video leaderboard (read 2026-10-10), MiniMax H3 (768p) ranks fourth with an Elo of 1137, behind Alibaba's Wan 3.0, Utopai X and ByteDance Seed's Dreamina Seedance 2.5. In its text-to-speech leaderboard, read the same day, MiniMax's highest-ranked model, Speech 2.8 HD, is 19th. On its language-model Intelligence Index, also read 2026-10-10, MiniMax-M3 scores 29, behind the latest models from Z AI, Kimi, Alibaba, Xiaomi and DeepSeek, and OpenRouter's top 20 models by weekly usage, as read 2026-10-10 (usage data through Oct 9, 2026), include none from MiniMax. The economics remain heavy: a loss of US$358.0 million in the half, research and development spending of US$296.9 million driven mainly by cloud services for training, and a July 2026 placing of new shares alongside a convertible bond issue, whose announcement says "prices of high-performance computing hardware have continued to rise amid industry-wide supply constraints in 2026". MiniMax is a fast-growing multimodal lab without a moat: its best ranking is in video, where it publishes the weights, and its language models trail Chinese rivals. |
| Chain position | A fabless designer sitting between its customers and the TSMC-centred manufacturing chain. The 2Q26 remarks describe 'deep design-technology co-optimization (DTCO) with TSMC' and close work with advanced-packaging partners to build 'high-performance ASICs across a broad range of very-large chip sizes using CoWoS and EMIB-T technologies' at nodes down to 2nm, and say MediaTek creates 'significant value beyond semiconductor by orchestrating other key components throughout the supply chain such as memory and substrate' -- a coordinator role as much as a design one. | Cipher develops single-tenant, powered data-centre campuses for hyperscale and AI tenants: Amazon at Black Pearl, Fluidstack (backstopped by Google) at Barber Lake, and an unnamed investment-grade hyperscaler at a third campus. Meanwhile it winds down bitcoin mining. | MiniMax is a model layer that rents its compute. The prospectus defines its AI infrastructure as mainly "computing services purchased from third-party cloud service providers, namely computing power, storage and network capacity that we rent from external cloud platforms instead of building and owning all the servers ourselves", and says its "upstream ecosystem includes major cloud service providers and infrastructure vendors". Downstream it sells model access to enterprises and developers through its Open Platform and runs consumer apps; 60.8% of revenue in the first half of 2026 came from outside Chinese mainland. |
| Products (share / barrier) |
|
|
|
| Long-horizon vote | +0.13 at weight 0.20 · swarm neutral Editorial prior, not backtested. | -0.06 at weight 0.20 · swarm bearish Editorial prior, not backtested. | -0.20 at weight 0.20 · swarm bullish Editorial prior, not backtested. |