"""Reproduces every number in 'The Economic Life of an Accelerator'.""" import numpy as np # Silicon Data neo-cloud rental indices, USD per GPU-hour, all dated 26 Oct 2026 P=dict(A100=1.56,H100=2.89,B200=6.63) F=dict(A100=312,H100=989,B200=2250) # BF16 dense TFLOPS BW=dict(A100=2.039,H100=3.35,B200=7.7) # TB/s M=dict(A100=80,H100=80,B200=180) # GB KW=dict(A100=6.5/8,H100=10.2/8,B200=14.3/8) # DGX system kW per GPU cs=['A100','H100','B200'] print('== Table 2 ==') for c in cs: print(c,f"rent {P[c]:.2f} $/PF-h {P[c]/(F[c]/1000):.2f} $/TBs-h {P[c]/BW[c]:.2f} $/GB-h {P[c]/M[c]:.3f} $/kW-h {P[c]/KW[c]:.2f}") y=np.log([P[c] for c in cs]) for name,X in [('throughput',F),('bandwidth',BW)]: x=np.log([X[c] for c in cs]); b,a=np.polyfit(x,y,1); r2=1-np.sum((y-(a+b*x))**2)/np.sum((y-y.mean())**2) print(f'line through log rent on log {name}: slope {b:.2f} R2 {r2:.3f}') for c in ['A100','H100']: fp=P['B200']/F['B200']*F[c]; bp=P['B200']/BW['B200']*BW[c] print(f'{c}: throughput parity {fp:.2f} (actual {P[c]:.2f}, {P[c]/fp-1:+.0%}); bandwidth parity {bp:.2f}') for prem in [0.004,0.058,0.121,0.162]: print(f'H200 premium {prem:.1%}: bandwidth elasticity {np.log(1+prem)/np.log(4.8/3.35):.2f}') # rent histories (Virtualized monthly series) def t(s): yy,m=s.split('-'); return int(yy)+(int(m)-0.5)/12 ser={'A100':[('2023-04',2.40),('2023-12',1.90),('2024-08',1.60),('2025-06',1.50),('2026-03',1.42),('2026-08',1.38),('2026-09',1.35),('2026-10',1.35),('2026-11',1.35)], 'H100':[('2024-10',3.20),('2025-04',2.85),('2025-10',2.75),('2026-03',2.65),('2026-08',2.55),('2026-09',2.49),('2026-10',2.49),('2026-11',2.69)], 'H200':[('2024-12',6.50),('2025-06',4.80),('2025-12',3.90),('2026-08',3.30),('2026-09',3.44),('2026-10',3.44),('2026-11',3.44)]} lam={} print('== decay fits ==') for c,pts in ser.items(): tt=np.array([t(s) for s,_ in pts]); pp=np.log([p for _,p in pts]); b,a=np.polyfit(tt,pp,1) r2=1-np.sum((pp-(a+b*tt))**2)/np.sum((pp-pp.mean())**2); lam[c]=-b print(f'{c}: {1-np.exp(b):.1%} a year, R2 {r2:.2f}, n {len(pts)}') print('H100 Apr 2023 to Oct 2024 annual decline', round(1-(3.20/8.00)**(1/1.5),3)) # operating cost colo=196.25/730; elec=0.0862; PUE=1.3; c_kw=colo+PUE*elec; u=0.7 print('== costs == opex per kW-h',round(c_kw,3)) for c in cs: print(c,f'opex {KW[c]*c_kw:.2f}/h, energy {KW[c]*PUE*elec:.2f}/h, rent/opex {P[c]/(KW[c]*c_kw):.1f} ({u*P[c]/(KW[c]*c_kw):.1f} at 70% use)') now=2026+9.85/12; launch=dict(A100=2020.37,H100=2022.75) print('== own-cost exit ages ==') for c in ['A100','H100']: for L in [lam[c],-np.log(0.75),-np.log(0.6)]: print(c,f'decline {1-np.exp(-L):.0%}: age {now-launch[c]+np.log(u*P[c]/(KW[c]*c_kw))/L:.1f}') # replacement test with levelised rents over five years r=0.10; n=60; v=(1+r)**(-1/12); tm=np.arange(n) def lev(R,L): w=v**tm; return np.sum(R*np.exp(-L*tm/12)*w)/np.sum(w) def capkwh(C=515410.43): return C/14.3*(r/(1-(1+r)**-5))/8760 def breakeven(c,LN,uu=u,C=515410.43): Rn=lev(P['B200'],LN); newnet=uu*Rn/KW['B200']-c_kw-capkwh(C) f=lev(1.0,lam[c]); return (newnet+c_kw)*KW[c]/uu/f, newnet print('== replacement test == capital charge',round(capkwh(),3)) for d in [0.0,0.094,0.129,0.25]: LN=-np.log(1-d); be={c:breakeven(c,LN) for c in ['A100','H100']} print(f'B200 rent decline {d:.1%}: B200 levelised net {be["A100"][1]:.2f}/kWh; A100 break-even {be["A100"][0]:.2f} ({P["A100"]/be["A100"][0]-1:+.0%}); H100 break-even {be["H100"][0]:.2f} ({P["H100"]/be["H100"][0]-1:+.0%})') for c in ['A100','H100']: f=lev(1.0,lam[c]); print(c,'levelised net now',round(u*P[c]*f/KW[c]-c_kw,2)) print('== Table 3, B200 rent falling 9.4% a year ==') for uu in [0.5,0.7,0.9]: print(uu,[ (round(breakeven('A100',-np.log(0.906),uu,C)[0],2),round(breakeven('H100',-np.log(0.906),uu,C)[0],2)) for C in [450000,515410.43,600000]]) print('interruptions per GPU-year',round((148+72)/16384*365/54,3))