Tether joins Tenstorrent $1.5B round despite QVAC cost data

Tether entered into Tenstorrent’s $1.5 billion Series E funding round on October 20, a move that extends its strategic push into the realm of AI hardware infrastructure, though the firm did not disclose the specific size of its stake. This participation follows an announcement made by Tenstorrent on October 19, which confirmed that the funding round had closed at a post-money valuation of $6.2 billion. Such a maneuver aligns with the broader vision held by CEO Paolo Ardoino regarding technology that is controlled by the user, yet fresh data emerging from QVAC, Tether’s local AI platform, suggests that the act of local inference does not automatically result in reduced costs for every single user.

QVAC’s Cost Analysis

On October 12, QVAC published a detailed cost study that compared the economics of local AI inference against the costs of accessing cloud APIs. The analysis takes the purchase price of a laptop costing $2,399 with 32 GB of memory and spreads that cost across a period of four years. This specific calculation results in a monthly hardware charge that comes to approximately $50. The study further estimates that a dense Qwen3.8-27B model processes roughly 55 million tokens during a standard working month consisting of 176 hours. The token mix used in this scenario consists of 94% input and 6% output, where tokens are defined as the units of text that a model both reads and generates.

Performance was measured using a 36 GB MacBook Pro M5 Max, which renders the comparison more illustrative than definitive. The original open-cloud comparison provided in the study places the break-even point at roughly 90 million tokens, a figure that exceeds the estimated capacity of the dense model for a single working month. However, current regional pricing shifts this threshold. The Singapore international schedule for Alibaba Cloud, which was checked on October 22, lists the Qwen3.8-27B model at $0.5 per million input tokens and $3 per million output tokens. These specific rates are applied before any promotions or discounts for caching are considered.

Calculating Break-Even

When applying the study’s 94/6 input-output mix, a blended price of $0.65 per million total tokens is produced. By dividing the $50 monthly hardware charge by this blended rate, the break-even point is placed at approximately 77 million tokens per month. At the estimated capacity of 55 million tokens, the corresponding bill for a cloud API would be $35.75. For those with lower usage, such as 22 million tokens, the API cost drops further to $14.3. If these specific usage levels and rates were to hold steady for four years, the bills avoided would not be enough to cover the purchase price of the laptop, and this excludes any operating costs or the remaining value of the hardware.

This comparison remains illustrative in nature. It operates on the assumption that local and hosted versions of the AI handle the same workload with an equal level of quality. It also excludes the cost of electricity and other general operating expenses. Furthermore, Alibaba lists different prices for Beijing and applies separate charges for inputs that are cached. A workload that qualifies for cheaper caching could lower the cloud bill even further. A different mix of input and output would also alter the blended rate. The estimate of 55 million tokens applies specifically to the model configuration of QVAC and the 176-hour window. While running for more hours increases the available processing time, the actual work must still fit within the capabilities of the model.

Hardware and Usage Scenarios

In a scenario involving QVAC’s faster Qwen3.6 35B mixture-of-experts, the system reaches an estimated 200 million tokens within the same working window. This higher capacity could accommodate an illustrative coding workload of 110 million tokens, which exceeds the estimate for the dense model. The study utilizes a Mac Studio for its example regarding coding-agent costs. Apple had announced the M5 Max desktop starting at $2,499 on September 22. Additionally, the sypherin benchmark repository reports 36.8 generated tokens per second for the dense Qwen3.8-27B on a 128 GB Ryzen AI MAX+395 system using FP4 weights. This community test employs a different configuration than the Mac used by QVAC, thereby providing a separate measurement with its own set of assumptions.

Those who already own adequate devices hold a much clearer advantage because the purchase cost is already a sunk cost. In such cases, local AI does not need to recover the full price of a new laptop. However, electricity costs still remain, and the device must possess sufficient speed and memory. For an individual buying a device solely for the purpose of occasional AI use, the hurdle is significantly higher. A fixed hardware cost persists even during quiet months, whereas API bills follow actual consumption. The example of 22 million tokens demonstrates that even substantial usage can leave a cheap hosted model well below the monthly allocation of a new laptop.

Users of sustained agents have a more plausible route toward achieving savings. This requires the use of a suitable model and a viable, faster configuration that can process enough comparable work. A fast model that requires more retries, or one that is unable to handle the task, would not match the economic result of a more capable service. The four-year allocation assumes that the buyer utilizes the hardware for that entire period. Actual cost recovery depends on the spending that the hardware replaces. For a device that is already needed for other types of work, allocating the entire purchase price to AI can overstate the additional cost of running a local model.

The argument presented by Ardoino includes benefits that extend beyond simple token prices. In the October 21 account of his October 7 AMA provided by QVAC, he advocates for AI that is owned by its users. He acknowledges that phones do not currently match the performance of frontier models and allows that the hardest scientific problems may continue to reside in large data centers. Local inference can keep downloaded models available while offline and reduce the dependence on the continued service of a provider. Keeping files and prompts on a local device can matter even when cloud tokens are cheaper. QVAC qualifies this privacy benefit by noting that connected tools can still send data to other locations.

Tenstorrent has stated that its new capital will support the deployment of Galaxy servers, as well as CPU and AI technology and commercial expansion. Neither of the financing announcements establishes a resulting reduction in user bills or a product integration with QVAC. Using the checked Singapore rates and the 176-hour dense-model example from QVAC, the break-even point still lies beyond the estimated capacity. A buyer in that specific scenario would need to place a value on privacy and control alongside any potential token savings.