AI capex is still accelerating, but the bottleneck has shifted from money to land, power, memory, and interconnects, so the trade is increasingly about which chip architecture captures the next spend. GOOGL, AMZN, MSFT, and META are on track to spend over $700 billion on data centers this year versus $375 billion in 2025, while their free cash flow is already cracking and they are funding capex with debt and, in Alphabet’s case, equity. Three competing answers are laid out: AMD/Talos on model-specific fixed-weight chips, Cerebras on wafer-scale chips, and NVDA on GPUs as financeable assets with longer usable lives. The closest near-term watchpoint is NVDA on August 26, plus whether the financing structures announced on August 10 actually survive scrutiny of GPU useful life, loan terms, and residual value.
GOOGL, AMZN, MSFT, and META are the core hyperscaler AI capex buyers, with spending compounding at over 60% annually since 2023 and expected to accelerate further.
Those four are on course to spend over $700 billion on data centers this year versus $375 billion in 2025, a 90% year-over-year jump.
AMZN free cash flow fell from $18 billion to negative $7.6 billion over the last 12 months, and AMZN raised 2026 capex from $200 billion to $220 billion because memory chips got pricier.
GOOGL added $15 billion to capex and posted negative free cash flow for the first time since its 2004 IPO.
META free cash flow is down 91%.
The hyperscalers are financing capex with debt and equity: AMZN sold $37 billion of bonds in Q1, GOOGL issued over $65 billion of bonds so far this year, and GOOGL announced an $85 billion equity raise in June, called the biggest in American corporate history.
Debt covered 9% of hyperscaler capex in 2024 and 32% today, showing the funding mix is shifting fast.
AI data-center buildouts are zero-sum because land and grid-connected power are limited; a major cluster can wait two to six years for a grid connection, and the next bottleneck shifts to cooling, compute density, network, and memory.
AMD’s answer is Talos: on August 6th AMD announced plans to acquire the 2023-founded startup, which etches AI model weights directly into metal layers so the chip no longer has to load weights from memory.
Talos’s approach trades general-purpose programmability for efficiency; AMD plans to pair Talos with Instinct GPUs and Helios racks so GPUs handle prompts and dynamic work while fixed-weight chips handle decode.
Talos chips are compared with NVDA’s Groq partnership: Groq LPUs use 500 MB of ultrafast on-chip memory, while Talos HC1 turns the model directly into hardware by making 1s into wires and 0s into broken paths.
Tenstorrent is framed as destroying the memory wall: a typical cloud AI provider serves Llama 3.1 8B at about 140 tokens per second per user, while Tenstorrent can hit 17,000 tokens per second, roughly 120x faster, with responses appearing instantly rather than line by line.
Tenstorrent uses a tiered manufacturing process with more than 100 physical layers so 98% of the chip stays fixed between models, only the top two metal layers change, and TSMC can add them and ship within 60 days after a new model is chosen.
Because the chip is on older 6nm, it has no HBM and simpler packaging; it is said to be up to 120x faster than a traditional GPU and roughly 20x cheaper, and Talus says its chips cost less than a penny per million tokens versus about 3.8 cents for an NVDA Black Widow on the same model.
Cerebras’s answer is the opposite of cutting wafers apart: WSE3 keeps the wafer intact, is about 8 inches a side, has 4 trillion transistors, 900,000 cores, 44 GB of SRAM, and 21 PB/s of bandwidth, about 2,600x NVDA Blackwell B200 memory bandwidth.
Cerebras said its backlog was $25.4 billion versus 2026 revenue guidance of $880 million to $890 million, with about 22% due in the next 2 years and another 43% in the following 2 years; hardware revenue was $82 million, cloud and services were $127 million, and gross margin was hit because demand exceeds capacity.
NVDA is pitching GPUs as an asset class: on August 10th it announced financing deals with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR for more than half a trillion dollars of AI infrastructure, and TensorRT-LLM doubled H100 inference performance while also working on Ampere, Ada Lovelace, and Blackwell.
The big NVDA risk is residual value: none of the financing deals are binding yet, no one has committed a dollar, NVDA may cover up to 25% of losses if equipment ends up worth less than the loan assumed, and Coreweave’s $2.6 billion GPU loan lasts 5 years against 3-year customer contracts.
The next NVDA earnings report is expected on August 26th, and the market still has to decide whether GPU useful life is 3 years or 5 years, because that determines whether GPUs can really be financed as assets.