adlrocha.substack.com/p/adlrocha-in-a-quest-to-becoming-ai?triedRedirect=true

GitHub's April 27, 2026 announcement says all Copilot plans remain in place and that each plan now includes a monthly allotment of GitHub AI Credits. It also says base plan pricing is not changing (for example, Copilot Pro remains $10/month and Pro+ remains $39/month). In other words, GitHub did not move to a pure pay-per-token model where everyone must pay for every token they use. Instead, it kept subscription plans and folded a monthly included credit balance into those plans, with additional usage available only after the included credits are exhausted and, for paid plans, if extra usage is allowed.

GitHub's billing docs say the same pattern existed even before the June 1 transition for premium requests: each plan included a monthly allowance, and extra usage was only billed if overages were enabled. So the statement that there are 'no more flat subscriptions' overstates the change.

NVIDIA's official Ampere documentation gives the GeForce RTX 3090 a memory bandwidth of 936 GB/sec per card. The article describes an 8× RTX 3090 configuration, so using the same 'aggregate bandwidth' logic applied elsewhere in the piece, the total is 8 × 936 GB/s = 7,488 GB/s (about 7.49 TB/s), not 936 GB/s.

That means the article is understating the 8-card system's aggregate bandwidth by a factor of eight. This matters because the same bandwidth figure is later used to compare the 8×3090 build against other multi-GPU and single-card systems.

NVIDIA's RTX PRO 6000 Blackwell datasheet lists 96 GB of memory and 1.8 TB/s of memory bandwidth per card. NVIDIA's Ampere documentation lists the RTX 3090 at 936 GB/s per card. For an 8× RTX 3090 machine, that works out to 7,488 GB/s (about 7.49 TB/s) in aggregate.

So while a single RTX PRO 6000 Blackwell has higher per-card bandwidth than a single RTX 3090, it is not faster in bandwidth than an entire 8-card RTX 3090 build. The comparison in the article mixes up per-card bandwidth with system-wide aggregate bandwidth.

tinygrad's official tinybox page lists the tinybox red v2 at 2,560 GB/s of aggregate GPU RAM bandwidth. NVIDIA's official Ampere documentation lists the RTX 3090 at 936 GB/s per card. Using the same aggregation approach the article uses elsewhere, an 8× RTX 3090 build totals 7,488 GB/s (about 7.49 TB/s).

So the tinybox red v2 does not 'punch well above' an 8×3090 build on bandwidth. It may be easier to buy and assemble, and it may compare favorably on other dimensions, but on raw aggregated memory bandwidth the 8×3090 setup is substantially higher.

Apple's official Mac Studio technical specifications list the M3 Ultra configuration at 819 GB/s memory bandwidth. Apple also describes the M3 Ultra Mac Studio as delivering over 800 GB/s of unified memory bandwidth in its newsroom announcement.

That means the table entry showing the M3 Ultra system at ~300 GB/s understates its memory bandwidth by well over a factor of two. The ~300 GB/s figure is not consistent with Apple's published specs for M3 Ultra.