Two AMD Radeon AI PRO R9700 cards, each carrying 32 GB of memory, were installed inside a workstation costing roughly $18,775 as tested.<\/p>\n
<\/p>\n The evaluation measured electricity draw, token throughput, and amortized hardware cost, then set those figures against several cloud subscription tiers for comparison.<\/p>\n Latest Videos FromTechRadar<\/p>\n How cloud pricing sets the bar<\/p>\n Cloud AI providers charge customers per million tokens generated, with prices ranging from $1.20 for GPT-5.6 Luna up to $30 for GPT-5.6 Sol.<\/p>\n Mid-tier models sit in between, with Claude Sonnet 5 at $10 and Claude Opus 5 at $25 per million output tokens generated.<\/p>\n You may like<\/p>\n The more expensive the cloud model a team would otherwise use, the sooner owned hardware pays for itself.<\/p>\n Testing both AMD cards together, the workstation<\/a> generated 156.2 tokens every second while serving eight simultaneous users during this test.<\/p>\n A speed technique called Multi-Token Prediction nearly doubled that figure, pushing throughput up to 320.2 tokens every second, with identical output quality.<\/p>\n At that 320.2 token-per-second speed, the machine only needs 3.5 hours of weekly use to beat GPT-5.6 Sol on cost, 4.2 hours to beat Claude Opus 5, and 8.8 hours to beat Gemini 3.1 Pro.<\/p>\n A team generating above 20 million tokens per month against GPT-5.6 Sol pricing gains real savings using this owned hardware setup.<\/p>\n What to read next<\/p>\n At that volume, running the workstation costs about $6,262 yearly in electricity and amortized hardware, against roughly $18,000 yearly in matching Sol fees.<\/p>\n That $11,738 yearly gap is the actual evidence behind the claim that heavy monthly usage makes AMD’s rig worthwhile.<\/p>\n If a company instead relies on GPT-5.6 Luna, priced at just $1.20 per million tokens, that math flips entirely in the other direction.<\/p>\n The workstation would then need 94.3 hours of weekly use just to match that far cheaper cloud subscription’s total cost.<\/p>\n Since a single week only contains 168 hours total, reaching that particular break-even point remains genuinely difficult without near constant, saturated usage.<\/p>\n Electricity itself was a minor factor throughout, since both cards together drew between 310 and 510 watts under sustained load conditions.<\/p>\n Where the economics tip in AMD’s favor<\/p>\n A smaller team producing only five million tokens monthly, priced against Gemini 3.1 Pro, would spend $6,262 yearly to displace just $720 in cloud costs, a clear loss.<\/p>\n That comparison shows the hardware only makes financial sense once usage climbs high enough to close a large yearly cost gap.<\/p>\n A single R9700 card handled an eight billion parameter AI model alone, processing 34.5 tokens every second without help from a second card.<\/p>\n Running just one card also lowers the effective break-even point further, since a lone card draws far less power under equivalent load.<\/p>\n That single card pulled between 221 and 283 watts depending on simultaneous user count, well under the 310 to 510 watts both cards drew together.<\/p>\n Smaller models therefore offer a second path into positive economics, letting lighter workloads justify a $1,880 single-card purchase long before a team can justify the full $3,760 dual-card upgrade.<\/p>\n Price alone still does not settle the question, since these locally run models measurably trail top cloud systems on complex reasoning benchmarks.<\/p>\n On one independent intelligence index, the 27 billion parameter AMD model scored 37 points against Google<\/a>‘s Gemini 3.1 Pro scoring 46 points instead.<\/p>\n Therefore, a team chasing the cheaper token count may be trading away real reasoning quality, not just cloud subscription fees.<\/p>\n Via Puget Systems<\/a><\/p>\n \n Follow TechRadar on Google News<\/a> and add us as a preferred source<\/a> to get our expert news, reviews, and opinion in your feeds.<\/p>\n","protected":false},"excerpt":{"rendered":"Two AMD cards cost $18,775 yet beat GPT-5.6 Sol within hours weeklyMulti-Token Prediction nearly doubled throughput to 320.2…\n","protected":false},"author":2,"featured_media":623976,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[61,60,80],"class_list":["post-623975","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology","tag-ie","tag-ireland","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts\/623975","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/comments?post=623975"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts\/623975\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/media\/623976"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/media?parent=623975"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/categories?post=623975"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/tags?post=623975"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}![\"Google](\"https:\/\/www.newsbeep.com\/ie\/wp-content\/uploads\/2026\/09\/1789354332_857_diM9tpwF2Lz85R8q85CT78.jpg\")
\n<\/p>\n